AMD Radeon 5700XT: Navi GPGPU Performance in OpenCL

What is “Navi”?

It is the code-name of the new AMD GPU, the first of the brand-new RDNA (Radeon DNA) GPU arch(itecture) – replacing the “Vega” that was the last of the GCN (graphics core next) arch(itecture). It is a mid-range GPU optimised for gaming thus not expected to set records, but GPUs today are used for many other tasks (mining, encoding, algorithm/compute acceleration, etc.) as well.

RDNA arch brings big changes from the various GCN revisions we’ve seen previously, but its first iteration here does not bring any major new features at least in the compute domain. Hopefully the next versions will bring tensor units (matrix multiplicators) and other accelerated instruction sets and so on.

See these other articles on GPGPU performance:

Hardware Specifications

We are comparing the middle-range Radeon with previous generation cards and competing architectures with a view to upgrading to a mid-range high performance design.

GPGPU Specifications AMD Radeon 5700XT (Navi) AMD Radeon VII (Vega2) nVidia Titan X (Pascal) AMD Radeon 56 (Vega1) Comments
Arch Chipset RDNA / Navi 10 GCN5.1 / Vega 20 Pascal / GP102 GCN5.0 / Vega 10 The first of the Navi chips.
Cores (CU) / Threads (SP) 40 / 2560 60 / 3840 28 / 3584 56 / 3584 Less CUs than Vega1 and same (64x) SP per CU.
SIMD per CU / Width 2 / 32 [2x] 4 / 16 4 / 16 Navi increases the SIMD width but decreases counts.
Wave/Warp Size 32 [1/2x] 64 32 64 Wave size is reduced to match nVidia.
Speed (Min-Turbo) 1.6 / 1.755 1.4 / 1.75 1.531 / 1.91 1.156 / 1.471 40% faster base and 20% turbo than Vega 1.
Power (TDP) 225W 295W 250W 210W Slightly higher TDP but nothing significant
ROP / TMU 64 / 160 64 / 240 96 / 224 64 / 224 ROPs are the same but we see ~30% less TMUs.
Shared Memory
64kB [+2x]
32kB 48kB / 96kB per SM 32kB We have 2x more shared memory allowing bigger kernels.
Constant Memory
4GB 8GB 64kB dedicated 4GB No dedicated constant memory but large.
Global Memory 8GB GDDR6 14Gt/s 256-bit 16GB HBM2 1Gt/s 4096-bit 12GB GDDR5X 10Gt/s 384-bit 8GB HBM2 900Gt/s 4096-bit Sadly no HBM this time but the faster but not very wide.
Memory Bandwidth (GB/s)
448GB/s [+9%] 1024GB/s 512GB/s 410GB/s Still bandwidth is 9% higher.
L1 Caches ? x40 16kB x60 48kB x28 16kB x56 L1 does not appear changed but unclear.
L2 Cache 4MB 4MB 3MB 4MB L2 has not changed.
Maximum Work-group Size
1024 / 1024 256 / 1024 1024 / 2048 per SM 256 / 1024 AMD has unlocked work-group sizes to 4x.
FP64/double ratio
1/16x 1/4x 1/32x 1/16x Ratio is same as consumer Vega1 rather than pro Vega2.
FP16/half ratio
2x 2x 1/64x 2x Ratio is the same throughout.

Processing Performance

We are testing both OpenCL performance using the latest SDK / libraries / drivers from both AMD and competition.

Results Interpretation: Higher values (GOPS, MB/s, etc.) mean better performance.

Environment: Windows 10 x64, latest AMD and nVidia drivers. Turbo / Boost was enabled on all configurations.

Processing Benchmarks AMD Radeon 5700XT (Navi) AMD Radeon VII (Vega2) nVidia Titan X (Pascal) AMD Radeon 56 (Vega1) Comments
GPGPU Arithmetic Benchmark Mandel FP16/Half (Mpix/s) 18,265 [-7%] 29,057 245 19,580 Navi starts well but cannot beat Vega1.
GPGPU Arithmetic Benchmark Mandel FP32/Single (Mpix/s) 11,863 [-13%] 17,991 17,870 13,550 Standard FP32 increases the gap to 13%.
GPGPU Arithmetic Benchmark Mandel FP64/Double (Mpix/s) 1,047 [-16%] 5,031 661 1,240 FP64 does not change much, Navi is 16% slower.
GPGPU Arithmetic Benchmark Mandel FP128/Quad (Mpix/s) 43 [-45%] 226 25 77 Emulated FP128 is hard on FP64 units and here Navi is almost 1/2 Vega1.
Starting up, Navi does not seem to be able to beat Vega1 in heavy vectorised compute loads with FP16 most efficient (almost parity) while complex FP128 is 2x slower.
GPGPU Crypto Benchmark Crypto AES-256 (GB/s) 51 [-25%] 91 42 67 Despite more bandwidth Navi is 25% slower than Vega1.
GPGPU Crypto Benchmark Crypto AES-128 (GB/s) 58 88
GPGPU Crypto Benchmark Crypto SHA2-256 (GB/s) 176 [+40%] 209 145 125 Navi shows its power here beating Vega1 by a huge 40%!
GPGPU Crypto Benchmark Crypto SHA1 (GB/s) 107 162
GPGPU Crypto Benchmark Crypto SHA2-512 (GB/s) 76 32
Despite more bandwidth of GDDR6, streaming algorithms work better on on “old” HBM2 thus Navi cannot beat Vega. But in pure integer compute algorithms like hashing, it is much faster by a significant amount which bodes well for the future.
GPGPU Finance Benchmark Black-Scholes float/FP32 (MOPT/s) 12,459 [+31%] 23,164 11,480 9,500 In this FP32 financial workload Navi is 30% faster than Vega1!
GPGPU Finance Benchmark Black-Scholes double/FP64 (MOPT/s) 7,272 1,370 1,880
GPGPU Finance Benchmark Binomial float/FP32 (kOPT/s) 850 [1/3x] 3,501 2,240 2,530 Binomial uses thread shared data thus stresses the memory system and here we have some optimisation to do.
GPGPU Finance Benchmark Binomial double/FP64 (kOPT/s) 789 129 164
GPGPU Finance Benchmark Monte-Carlo float/FP32 (kOPT/s) 5,027 [+30%] 6,249 5,350 3,840 Monte-Carlo also uses thread shared data but read-only thus reducing modify pressure – here Navi is again 30% faster.
GPGPU Finance Benchmark Monte-Carlo double/FP64 (kOPT/s) 1,676 294 472
For financial FP32 workloads, Navi is ~30% faster than Vega1 – a pretty good improvement – though it naturally cannot compete with Vega2 due to consumer multiplier (1/16x). Crypto-currencies fans will love the Navi.
GPGPU Science Benchmark SGEMM (GFLOPS) float/FP32 5,165 [+2%] 6,634 6,073 5,066 GEMM can only bring a measly 2% improvement over Vega1.
GPGPU Science Benchmark DGEMM (GFLOPS) double/FP64 2,339 340 620
GPGPU Science Benchmark SFFT (GFLOPS) float/FP32 376 [+2%] 643 235 369 FFT loves HBM but Navi is still 2% faster.
GPGPU Science Benchmark DFFT (GFLOPS) double/FP64 365 207 175
GPGPU Science Benchmark SNBODY (GFLOPS) float/FP32 4,534 [-6%] 6,846 5,720 4,840 Navi can’t manage as well in N-Body and ends up 6% slower.
GPGPU Science Benchmark DNBODY (GFLOPS) double/FP64 1,752 275 447
The scientific scores don’t show the same improvement as the financial ones likely due to heavy use of shared memory with Navi just matching Vega1. Perhaps the larger shared memory can allow us to use larger workgroups.
GPGPU Image Processing Blur (3×3) Filter single/FP32 (MPix/s) 8,674 [1/2.1x] 25,418 18,410 19,130 In this 3×3 convolution algorithm, Navi is 1/2x the speed of Vega1.
GPGPU Image Processing Sharpen (5×5) Filter single/FP32 (MPix/s) 1,734 [1/3x] 5,275 5,000 4,340 Same algorithm but more shared data makes Navi even slower.
GPGPU Image Processing Motion Blur (7×7) Filter single/FP32 (MPix/s) 1,802 [1/2.5x] 5,510 5,080 4,450 With even more data the gap remains at 1/2.5x.
GPGPU Image Processing Edge Detection (2*5×5) Sobel Filter single/FP32 (MPix/s) 1,723 [1/2.5x] 5,273 4,800 4,300 Still convolution but with 2 filters – same 1/2.5x performance.
GPGPU Image Processing Noise Removal (5×5) Median Filter single/FP32 (MPix/s) 48.44 [=] 92.53 37 48 Different algorithm allows Navi to tie with Vega1.
GPGPU Image Processing Oil Painting Quantise Filter single/FP32 (MPix/s) 97.34 [+2.5x] 57.66 12.7 38 Without major processing, this filter performs well on Navi.
GPGPU Image Processing Diffusion Randomise (XorShift) Filter single/FP32 (MPix/s) 32,050 [+1.5x] 47,349 19,480 20,880 This algorithm is 64-bit integer heavy and Navi is 50% faster than Vega1.
GPGPU Image Processing Marbling Perlin Noise 2D Filter single/FP32 (MPix/s) 9,516 [+1.6x] 7,708 305 6,000 One of the most complex and largest filters, Navi is again 50% faster.
For image processing using FP32 precision, Navi goes from 1/2.5x Vega1 performance (convolution) to 50% faster (complex algorithms with integer processing). It seems some optimisations are needed for the convolution algorithms.

Memory Performance

We are testing both OpenCL performance using the latest SDK / libraries / drivers from AMD and competition.

Results Interpretation: For bandwidth tests (MB/s, etc.) high values mean better performance, for latency tests (ns, etc.) low values mean better performance.

Environment: Windows 10 x64, latest AMD and nVidia. drivers. Turbo / Boost was enabled on all configurations.

Memory Benchmarks AMD Radeon 5700X (Navi) AMD Radeon VII (Vega2) nVidia Titan X (Pascal) AMD Radeon 56 (Vega1) Comments
GPGPU Memory Bandwidth Internal Memory Bandwidth (GB/s) 376 [+13%] 627 356 333 Navi’s GDDR6 manages 13% more bandwidth than Vega1.
GPGPU Memory Bandwidth Upload Bandwidth (GB/s) 21.56 [+77%] 12.37 11.4 12.18 PCIe 4.0 brings almost 80% more bandwidth
GPGPU Memory Bandwidth Download Bandwidth (GB/s) 22.28 [+84%] 12.95 12.2 12.08 Again almost 2x more bandwidth.
Navi’s PCIe 4.0 interface (on 500-series motherboards) brings as expected almost 2x more upload/download bandwidth while its high-clocked GDDR6 manages just over 10% higher bandwidth over HBM2.
GPGPU Memory Latency Global (In-Page Random Access) Latency (ns) 276 [+11%] 202 201 247 Navi’s GDDR6 brings slight latency increase (+10%)
GPGPU Memory Latency Global (Full Range Random Access) Latency (ns) 341 286 353
GPGPU Memory Latency Global (Sequential Access) Latency (ns) 89.8 115
GPGPU Memory Latency Constant Memory (In-Page Random Access) Latency (ns) 117 237
GPGPU Memory Latency Shared Memory (In-Page Random Access) Latency (ns) 18.7 55
GPGPU Memory Latency Texture (In-Page Random Access) Latency (ns) 195 193
GPGPU Memory Latency Texture (Full Range Random Access) Latency (ns) 282 301
GPGPU Memory Latency Texture (Sequential Access) Latency (ns) 87.6 80
Not unexpected, GDDR6′ latencies are higher than HBM2 although not by as much as we were fearing.

SiSoftware Official Ranker Scores

Final Thoughts / Conclusions

“Navi” is an interesting chip to be sure and perhaps more was expected of it; as always the drivers are the weak link and it is hard to determine which issues will be fixed driver-side and which will need to be optimised in compute kernels.

Thus performance-wise it oscillates between 1/2x and 50% Vega1 performance depending on algorithm, with compute-heavy algorithms (especially crypto-currencies) doing best and shared/local memory heavy algorithms doing worst. The 2x bigger shared memory (64kB vs 32) in conjunction with the larger work-group (1024 vs 256 by default) sizes do present future optimisation opportunities. AMD has also reduced the warp/wave size to match nVidia – a historic change.

Memory wise, the cost-cutting change from HBM2 to even high-speed GDDR6 does bring more bandwidth but naturally higher latencies – but PCIe 4.0 doubles upload/download bandwidths which will become much more important on higher capacity (16GB+) cards in the future.

Overall it is hard to recommend it for compute workloads unless the particular algorithm (crypto, financial) does well on Navi, otherwise the much older Vega1 56/64 offer better performance/cost ratio especially today. However, as drivers mature and implementations are optimised for it, Navi is likely to start to perform better.

We are looking forward to the next iterations of Navi, especially the rumoured “big Navi” version optimised for compute…

AMD Radeon VII: Vega2 GPGPU Performance in OpenCL

What is “Vega2”?

It is the code-name of the updated “Vega” GPU arch(itecture), the last of the GCN (graphics core next) arch (version 5.1) shrinked to 7nm before being replaced by the forthcoming “Navi”. Originally for the professional/workstation high-end market, “Vega2″/”big Vega” designed for compute (scientific, machine learning, etc.) workloads was pressed into service to battle the latest 2000-series “Turing”/RTX competition.

As a result it contains many high-end features not normally found on consumer cards:

  • 1/4 FP64 rate (instead of 1/16 or worse)
  • 16GB HBM2 memory (instead of 8-12)
  • 4096-bit HBM2 memory 1TB/s bandwidth (instead of 400-500)
  • Int8/Int4 support for AI/ML workloads
  • PCIe 4.0 capable but not enabled at this time

See these other articles on GPGPU performance:

Hardware Specifications

We are comparing the middle-range Radeon with previous generation cards and competing architectures with a view to upgrading to a mid-range high performance design.

GPGPU Specifications AMD Radeon VII (Vega2) nVidia Titan V (Volta) nVidia Titan X (Pascal) AMD Vega 56 (Vega1) Comments
Arch Chipset Vega 20 / GCN 5.1 GV100 / 7.0 GP102 / 6.1 Vega 10 / GCN 5.0 A minor revision of Vega1.
Cores (CU) / Threads (SP) 60 / 3840 80 / 5120 28 / 3584 56 / 3584 More CUs than normal Vega but not 64.
SIMD per CU / Width 4 / 16 n/a n/a 4 / 16 Naturally same SIMD count and width
Wave/Warp Size 64 32 32 75 Wave size has always been 2x nVidia.
Speed (Min-Turbo) 1.4 – 1.750 [+21%] (135-1455) 1.531 (139-1910) 1.156 – 1.471 Base clock is ~20% higher and turbo
Power (TDP) 300W [+42%] 300W 250W 210W TDP has gone up by 40%.
ROP / TMU 64 / 256 96 / 320 96 / 224 64 / ROPs and TMUs unchanged
Shared Memory
32kB 48 / 96 kB 48 / 96kB 32kB No shared memory change.
Constant Memory
8GB 64kB 64kB 4GB No dedicated constant memory but large.
Global Memory 16GB HBM2 2Gbps 4096-bit 12GB HBM2 2x850Mbps 3072-bit 12GB GDDR5X 10Gbps 384-bit 8GB HBM2 1.89Gbps 2048-bit 2x as big and 2x as wide HBM a huge improvement.
Memory Bandwidth (GB/s)
1000 [+2.4x] 652 512 410 Still bandwidth is 9% higher.
L1 Caches 16kB x 60 96kB x 80 48kB x 28 16kB x 56 L1 has not changed.
L2 Cache 4MB 4.5MB 3MB 4MB L2 has not changed.
Maximum Work-group Size
256 / 1024 1024 / 2048 1024 / 2048 256 / 1024 Same work-group sizes.
FP64/double ratio
1/4x 1/2x 1/32x 1/16x Ratio is 4x better than Vega1.
FP16/half ratio
2x 2x 1/64x 2x Ratio is the same throughout.

Processing Performance

We are testing both OpenCL performance using the latest SDK / libraries / drivers from both AMD and competition.

Results Interpretation: Higher values (GOPS, MB/s, etc.) mean better performance.

Environment: Windows 10 x64, latest AMD and nVidia drivers. Turbo / Boost was enabled on all configurations.

Processing Benchmarks AMD Radeon VII (Vega2) nVidia Titan V (Volta) nVidia Titan X (Pascal) AMD Vega 56 (Vega1) Comments
GPGPU Arithmetic Benchmark Mandel FP16/Half (Mpix/s) 29,057 [+48%] 33,860 245 19,580 Vega2 starts strong with a 48% lead over Vega1 and almost catching Volta.
GPGPU Arithmetic Benchmark Mandel FP32/Single (Mpix/s) 18,340 [+35%] 22,680 17,870 13,550 Good improvement here +35% over Vega1 again close to Volta.
GPGPU Arithmetic Benchmark Mandel FP64/Double (Mpix/s) 5,377 [+4.3x] 11,000 661 1,240 1/4 FP64 rate makes it over four (4x) times faster than Vega1.
GPGPU Arithmetic Benchmark Mandel FP128/Quad (Mpix/s) 234 [+3x] 458 25.77 77 Similar to above, Vega2 is over three (3x) faster.
Vega2 looks about 35-50% faster than Vega1 in FP32/FP16 and 3-4x faster in FP64 due to its 1/4 FP64 rate. It won’t beat real workstation cards with 1/2 FP64 rate through thus that Titan has nothing to worry about.
GPGPU Crypto Benchmark Crypto AES-256 (GB/s) 91 [+36%] 70 42 67 The fast HBM2 memory allows it to beat even Volta not just Vega1.
GPGPU Crypto Benchmark Crypto AES-128 (GB/s) 93 58 88
GPGPU Crypto Benchmark Crypto SHA2-256 (GB/s) 209 [+67%] 245 145 125 Vega2 is a huge 70% faster in integer/crypto workloads.
GPGPU Crypto Benchmark Crypto SHA1 (GB/s) 129 107 162
GPGPU Crypto Benchmark Crypto SHA2-512 (GB/s) 176 76 32
Vega2 increases its lead in integer workloads even streaming ones no doubt due to its very fast HBM2 memory making it the crypto-king of the hill though its cost may be an issue.
GPGPU Finance Benchmark Black-Scholes float/FP32 (MOPT/s) 23,164 [+2.3x] 18,570 11,480 9,500 Vega2 is over 2x faster than Vega1 also beating Volta.
GPGPU Finance Benchmark Black-Scholes double/FP64 (MOPT/s) 7,272 [+3.84x] 8,400 1,370 1,880 In FP64 its almost 4x faster just below Volta!
GPGPU Finance Benchmark Binomial float/FP32 (kOPT/s) 3,501 [+38%] 4,200 2,240 2,530 Binomial uses thread shared data thus stresses the memory system Vega2 is still 40% faster.
GPGPU Finance Benchmark Binomial double/FP64 (kOPT/s) 789 [+4.8x] 2,000 129 164 With FP64 we’re almost 5x faster than Vega1.
GPGPU Finance Benchmark Monte-Carlo float/FP32 (kOPT/s) 6,249 [+62%] 11,920 5,350 3,840 Monte-Carlo also uses thread shared data but read-only thus reducing modify pressure – here Vega2 is 60% faster.
GPGPU Finance Benchmark Monte-Carlo double/FP64 (kOPT/s) 1,676 [+3.55x] 4,440 294 472 With FP64 we’re over 3.5x faster.
For financial FP32 workloads, Vega2 is 40-60% faster than Vega1 a decent improvement; naturally in FP64 it’s 4-5x times faster thus a significant upgrade for algorithms that require such precision.
GPGPU Science Benchmark SGEMM (GFLOPS) float/FP32 6,634 [+30%] 11,000 6,073 5,066 GEMM still brings a 30% improvement over Vega1.
GPGPU Science Benchmark DGEMM (GFLOPS) double/FP64 2,339 [+3.77x] 3,830 340 620 But DGEMM is almost 4x faster.
GPGPU Science Benchmark SFFT (GFLOPS) float/FP32 643 [+74%] 617 235 369 FFT loves HBM thus Vega2 is 75% faster.
GPGPU Science Benchmark DFFT (GFLOPS) double/FP64 365 [+2.1x] 280 207 175 DFFT is tough but Vega2 is still twice as fast.
GPGPU Science Benchmark SNBODY (GFLOPS) float/FP32 6,846 [+41%] 7,790 5,720 4,840 In N-Body physics Vega2 is 40% faster.
GPGPU Science Benchmark DNBODY (GFLOPS) double/FP64 1,752 [+3.9x] 4,270 275 447 And in FP64 physics Vega2 is almost 4x faster.
The scientific scores show a similar improvement, with FP32 30-40% better but FP64 a whopping four (4x) faster than Vega1 and, in some algorithms, matching the hugely expensive Volta.
GPGPU Image Processing Blur (3×3) Filter single/FP32 (MPix/s) 25,418 [+32%] 26,790 18,410 19,130 In this 3×3 convolution algorithm, Vega2 is 32% faster than Vega1
GPGPU Image Processing Sharpen (5×5) Filter single/FP32 (MPix/s) 5,275 [+21%] 9,295 5,000 4,340 Same algorithm but more shared data reduces the lead to 21%.
GPGPU Image Processing Motion Blur (7×7) Filter single/FP32 (MPix/s) 5,510 [+24%] 9,428 5,080 4,450 With even more data the gap remains constant.
GPGPU Image Processing Edge Detection (2*5×5) Sobel Filter single/FP32 (MPix/s) 5,273 [+23%] 9,079 4,800 4,300 Still convolution but with 2 filters – similar 23% faster.
GPGPU Image Processing Noise Removal (5×5) Median Filter single/FP32 (MPix/s) 92 [+91%] 112 37 48 Different algorithm makes Vega2 almost 2x faster than Vega1.
GPGPU Image Processing Oil Painting Quantise Filter single/FP32 (MPix/s) 57 [+50%] 42 12.7 38 Without major processing, this filter is 50% faster on Vega2.
GPGPU Image Processing Diffusion Randomise (XorShift) Filter single/FP32 (MPix/s) 47,349 [+2.3x] 24,370 19,480 20,880 This algorithm is 64-bit integer heavy and Vega2 flies 2x faster than Vega1.
GPGPU Image Processing Marbling Perlin Noise 2D Filter single/FP32 (MPix/s) 7,708 [+28%] 8,460 305 6,000 One of the most complex and largest filters, Vega2 is 28% faster.
For image processing using FP32 precision, Vega goes from 21% to 2x faster, overall a decent improvement if you are processing a large number of images. In many filters it beats the far more expensive Volta competition.

Memory Performance

We are testing both OpenCL performance using the latest SDK / libraries / drivers from AMD and competition.

Results Interpretation: For bandwidth tests (MB/s, etc.) high values mean better performance, for latency tests (ns, etc.) low values mean better performance.

Environment: Windows 10 x64, latest AMD and nVidia. drivers. Turbo / Boost was enabled on all configurations.

Memory Benchmarks AMD Radeon VII (Vega2) nVidia Titan V (Volta) nVidia Titan X (Pascal) AMD Vega 56 (Vega1) Comments
GPGPU Memory Bandwidth Internal Memory Bandwidth (GB/s) 627 [+88%] 536 356 333 Vega2’s wide HBM2 is almost 2x faster as expected.
GPGPU Memory Bandwidth Upload Bandwidth (GB/s) 12.37 [+2%] 11.47 11.4 12.18 Using PCIe 3.0 similar upload bandwidth.
GPGPU Memory Bandwidth Download Bandwidth (GB/s) 12.95 [+7%] 12.27 12.2 12.08 Again similar bandwidth.
Vega2 benefits greatly from its very wide HBM2 memory (4096-bit) which provides almost 2x real bandwidth as expected. But while PCIe 4.0 capable for now it has to make do with 3.0 and thus same upload/download bandwith. Here’s hoping for a BIOS update once new motherboards come out.
GPGPU Memory Latency Global (In-Page Random Access) Latency (ns) 202 [-19%] 180 201 247 The higher clock allows Vega2 a 20% latency reduction.
GPGPU Memory Latency Global (Full Range Random Access) Latency (ns) 341 [-4%] 311 286 353 Full range is only 4% faster.
GPGPU Memory Latency Global (Sequential Access) Latency (ns) 53.4 89.8 115
GPGPU Memory Latency Constant Memory (In-Page Random Access) Latency (ns) 75.4 117 237
GPGPU Memory Latency Shared Memory (In-Page Random Access) Latency (ns) 18.1 18.7 55
GPGPU Memory Latency Texture (In-Page Random Access) Latency (ns) 212 195 193
GPGPU Memory Latency Texture (Full Range Random Access) Latency (ns) 344 282 301
GPGPU Memory Latency Texture (Sequential Access) Latency (ns) 88.5 87.6 80
Not unexpected, GDDR6′ latencies are higher than HBM2 although not by as much as we were fearing.

SiSoftware Official Ranker Scores

Final Thoughts / Conclusions

Vega2 (“BigVega”) is a big improvement over normal Vega1 and its workstation-class pedigree shows. For FP16/Fp32 workloads though the 30-40% performance improvement may not be worth it considering the much higher price: naturally FP64 performance is almost 4x due to 1/4 FP64 rate though not as good at professional cards with 1/2 rate or Titan competition with similar 1/2 rate.

While the GCN core (rev 5.1) has seen internal updates, there is nothing new that can be supported/optimised for in the compute land thus any code working well on Vega1 should work just as well on Vega2.

The 16GB HBM2 wide memory also helps big workloads with 2x higher bandwidth and also lower latency due to higher clock. For some workloads this alone makes it a definite buy when competition stops at 12GB.

Unfortunately the card has had a limited release at a relatively high price thus value/price ratio depends entirely on your workload – if FP64 with large datasets then it is very much worth it; if FP32/FP16 with datasets that fit in standard 8GB memory then the older Vega1 is much better value and you can even get 2 for the price of the Vega2.

For revolutionary change we need to wait for Navi and its brand new RDNA (Radeon DNA) arch(itecture)…

SiSoftware Sandra 20/20 (2020) Released!

FOR IMMEDIATE RELEASE

Contact: Press Office

SiSoftware Sandra 20/20 (2020) Released:
Brand-new benchmarks (AI/ML), hardware support

Updates: RTMa (30.16, August 8th).

London, UK, July 18th, 2019 – We are pleased to announce the launch of SiSoftware Sandra 20/20 (2020), the latest version of our award-winning utility, which includes remote analysis, benchmarking and diagnostic features for PCs, servers, mobile devices and networks.

It adds two Neural Networks AI/ML (Artificial Intelligence/Machine Learning) benchmarks for both CPU and GP (GPU) to measure both CNN (Convolution Neural Network) & RNN (Recurrent Neural Networks) performance on modern hardware.

It also adds hardware support and optimisations for brand-new CPU architectures (AMD Ryzen 2 (3000 series); Intel IceLake, CometLake) not forgetting GPGPU architectures across the various interfaces (CUDA, OpenCL, DirectX ComputeShader, OpenGL Compute).

As SiSoftware operates a “just-in-time” release cycle, some features were introduced in Sandra 2017 service packs: in Sandra Titanium they have been updated and enhanced based on all the feedback received.

Operating System Module

Broad Operating System Support

All current versions supported: Windows 10, 8.1*, 8*, 7*; Server 2019, 2016, 2012/R2 and 2008/R2*

Brand new AI/ML benchmarks featuring both CNN & RNN networks testing both inference/forward and training/back-propagation performance.

Processor Neural Networks (AI/ML)

A combined performance index of CNN (inference/forward & training) & RNN (inference/forward & training) for all precisions (single/FP32, double/FP64 floating-point) and instruction sets (AVX512, AVX2/FMA, AVX, SSE4, SSE2, RTM/HLE with NUMA and large-page support)

Ranker: Processor Neural Networks (Normal/Single Precision)
Ranker: Processor Neural Networks (High/Double Precision)

GP (GPU) Neural Networks (AI/ML)

A combined performance index of CNN (inference/forward & training) & RNN (inference/forward & training) for all precisions (half/FP16, single/FP32 floating-point) and platforms (CUDA, OpenCL, DirectX Compute)

GP (GPU) Neural Networks (Normal/Single Precision)
GP (GPU) Neural Networks (Low/Half Precision)

CNN (Convolution Neural Network) Architecture

Detailed document on the CNN architecture, data-sets and results that underpin our choices for the new benchmarks.

The new Neural Networks (AI/ML) Benchmarks: CNN Architecture

RNN (Recurrent Neural Network) Architecture

Detailed document on the RNN architecture, data-sets and results that underpin our choices for the new benchmarks.

The new Neural Networks (AI/ML) Benchmarks: RNN Architecture

Major changes

  • All connections to website engines (Ranker, Information, Price) are now secured by SSL through HTTP.
  • Sandra client (management console) is now installed as native 64-bit (on x64 and arm64) and thus needs 64-bit Access components (2016, 2013, 2010, etc.) or SQL Server (2017, 2016, 2014, etc) for its database.

Key features of Sandra 20/20

  • 4 native architectures support (x86, x64, ARM64** – Windows; ARM, ARM64, x86, x64 – Android)
  • Huge official hardware support through technology partners (AMD/ATI, nVidia, Intel).
  • 4 native (GP)GPU/APU platforms support (OpenCL 2.1+, CUDA 10.1+, DirectX Compute Shader 11/10+, OpenGL Compute 4.5+, Vulkan 1.0+).
  • 4 native Graphics platforms support (DirectX 11.x/10.x, OpenGL 4.0+, Vulkan 1.0+).
  • 9 language versions (English, German, French, Italian, Spanish, Japanese, Chinese (Traditional, Simplified), Russian) in a single installer.
  • Enhanced Sandra Lite (Eval) version (free for personal/educational use, evaluation for other uses)

Articles & Benchmarks

For more details, please see the following articles:

Purchasing

For more details, and to purchase the commercial versions, please click here.

Updating or Upgrading

To update your existing commercial version, please click here.

Downloading

For more details, and to download the Lite (Evaluation) version, please click here.

Reviewers and Editors

For your free review copies, please contact us.

About SiSoftware

SiSoftware, founded in 1995, is one of the leading providers of computer analysis, diagnostic and benchmarking software. The flagship product, known as “SANDRA”, was launched in 1997 and has become one of the most widely used products in its field. Many worldwide IT publications, magazines and review sites use SANDRA to analyse the performance of today’s computers. Thousands on-line reviews of computer hardware that use SANDRA are catalogued on our website alone.

Since launch, SiSoftware has always been at the forefront of the technology arena, being among the first providers of benchmarks that show the power of emerging new technologies such as multi-core, GPGPU, OpenCL, OpenGL, DirectCompute, x64, ARM64, ARM, NUMA, SMT (Hyper-Threading), SMP (multi-threading), AVX512, AVX2/FMA3, AVX, NEON/2, SSE4.2/4, SSSE3, SSE2, SSE, Java and .NET.

SiSoftware is located in London, UK. For more information, please visit www.sisoftware.net, www.sisoftware.eu, or www.sisoftware.co.uk

AMD Ryzen2 3700X Review & Benchmarks – CPU 8-core/16-thread Performance

What is “Ryzen2” ZEN2?

AMD’s Zen2 (“Matisse”) is the “true” 2nd generation ZEN core on 7nm process shrink while the previous ZEN+ (“Pinnacle Ridge”) core was just an optimisation of the original ZEN (“Summit Ridge”) core that while socket compatible it introduces many design improvements over both previous cores. An APU version (with integrated “Navi” graphics) is scheduled to be launched later.

While new chipsets (500 series) will also be introduced and required to support some new features (PCIe 4.0), with an BIOS/firmware update older boards may support them thus allowing upgrades to existing systems adding more cores and thus performance. [Note: older boards will not be enabled for PCIe 4.0 after all]

The list of changes vs. previous ZEN/ZEN+ is extensive thus performance delta is likely to be very different also:

  • Built around “chiplets” of up to 2 CCX (“core complexes”) each of 4C/8T and 8MB L3 cache (7nm)
  • Central I/O hub with memory controller(s) and PCIe 4.0 bridges connected through IF (“Infinity Fabric”) (12nm)
  • Up to 2 chiplets on desktop platform thus up to 2x2x4C (16C/32T 3950X) (same amount as old ThreadRipper 1950X/2950X)
  • 2x larger L3 cache per CCX thus up to 2x2x16MB (64MB) L3 cache (3900X+)
  • 24 PCIe 4.0 lanes (2x higher transfer rate over PCIe 3.0)
  • 2x DDR4 memory controllers up to 4266Mt/s

To upgrade from Ryzen+/Ryzen1 or not?

Micro-architecturally there are more changes that should improve performance:

  • 256-bit (single-op) SIMD units 2x Fmacs (fixing a major deficiency in ZEN/ZEN+ cores)
  • TLB (2nd level) increased (should help out-of-page access latencies that are somewhat high on ZEN/ZEN+)
  • Memory latencies claim to be reduced through higher-speed memory (note all requests go through IF to Central I/O hub with memory controllers)
  • Load/Store 32bytes/cycle (2x ZEN/ZEN+) to keep up with the 256-bit SIMD units (L1D bandwidth should be 2x)
  • L3 cache is 2x ZEN/ZEN+ but higher latency (cache is exclusive)
  • Infinity Fabric is 512-bit (2x ZEN/ZEN+) and can run 1x or 1/2x vs. DRAM clock (when higher than 3733Mt/s)
  • AMD processors have thankfully not been affected by most of the vulnerabilities bar two (BTI/”Spectre”, SSB/”Spectre v4″) that have now been addressed in hardware.
  • HWM-P (hardware performance state management) transitions latencies reduced (ACPI/CPPCv2)

In this article we test CPU core performance; please see our other articles on:

Hardware Specifications

We are comparing the middle-of-the-range Ryzen2 (3700X) with previous generation Ryzen+ (2700X) and competing architectures with a view to upgrading to a mid-range high performance design.

CPU Specifications AMD Ryzen 9 3900X (Matisse)
AMD Ryzen 7 3700X (Matisse) AMD Ryzen 7 2700X (Pinnacle Ridge) Intel i9 9900K (Coffeelake-R) Intel i9 7900X (Skylake-X) Comments
Cores (CU) / Threads (SP) 12C / 24T 8C / 16T 8C / 16T 8C / 16T 10C / 20T Core counts remain the same.
Topology 2 chiplets, each 2 CCX, each 3 cores (1 disabled) (12C) 1 chiplet, 2 CCX, each 4 cores (8C) 2 CCX, each 4 cores (8C) Monolithic die Monolithic die 1 chiplet+1 sio rather than 1 die
Speed (Min / Max / Turbo) 3.8 / 4.6GHz 3.6 / 4.4GHz 3.7 / 4.2GHz 3.6 / 5GHz 3.3 / 4.3GHz 3700x base clock is lower than 2700x but turbo is higher
Power (TDP / Turbo) 105 / 135W 65 / 90W 105 / 135W 95 / 135W 140 / 308W TDP has been greatly reduced vs. ZEN+
L1D / L1I Caches 12x 32kB 8-way / 12x 32kB 8-way 8x 32kB 8-way / 8x 32kB 8-way 8x 32kB 8-way / 8x 64kB 4-way 8x 32kB 8-way / 8x 32kB 8-way 10x 32kB 8-way / 10x 32kB 8-way L1I has been halved but better no. ways
L2 Caches 12x 512kB (6MB) 8-way 8x 512kB (4MB) 8-way 8x 512kB (4MB) 8-way 8x 256kB (2MB) 16-way 10x 1MB (10MB) 16-way No changes to L2
L3 Caches 2x2x 16MB (64MB) 16-way 2x 16MB (32MB) 16-way 2x 8MB (16MB) 16-way 16MB 16-way 13.75MB 11-way L3 is 2x ZEN+
Mitigations for Vulnerabilities BTI/”Spectre”, SSB/”Spectre v4″ hardware BTI/”Spectre”, SSB/”Spectre v4″ hardware BTI/”Spectre”, SSB/”Spectre v4″ software/firmware RDCL/”Meltdown”, L1TF hardware, BTI/”Spectre”, MDS/”Zombieload”, software/firmware RDCL/”Meltdown” , L1TF, BTI/”Spectre”, MDS/”Zombieload”, all software/firmware Ryzen2 addresses the remaining 2 vulnerabilities while Intel was forced to add MDS to its long list…
Microcode MU-8F7100-11 MU-8F7100-11 MU-8F0802-04 MU-069E0C-9E MU-065504-49 The latest microcodes included in the respective BIOS/Windows have been loaded.
SIMD Units 256-bit AVX/FMA3/AVX2 256-bit AVX/FMA3/AVX2 128bit AVX/FMA3/AVX2 256-bit AVX/FMA3/AVX2 512-bit AVX512 ZEN2 SIMD units are 2x wider than ZEN+

Native Performance

We are testing native arithmetic, SIMD and cryptography performance using the highest performing instruction sets (AVX2, FMA3, AVX, etc.). Ryzen2 supports all modern instruction sets including AVX2, FMA3 and even more like SHA HWA but not AVX-512.

Results Interpretation: Higher values (GOPS, MB/s, etc.) mean better performance.

Environment: Windows 10 x64, latest AMD and Intel drivers. 2MB “large pages” were enabled and in use. Turbo / Boost was enabled on all configurations. All mitigations for vulnerabilities (Meltdown, Spectre, L1TF, MDS, etc.) were enabled as per Windows default where applicable.

Native Benchmarks AMD Ryzen 7 3700X (Matisse)
AMD Ryzen 7 2700X (Pinnacle Ridge)
Intel i9 9900K (Coffeelake-R)
Intel i9 7900X (Skylake-X)
Comments
CPU Arithmetic Benchmark Native Dhrystone Integer (GIPS) 336 [=] 334 400 485 We start with no improvement over ZEN+
CPU Arithmetic Benchmark Native Dhrystone Long (GIPS) 339 [=] 335 393 485 With a 64-bit integer workload nothing much changes.
CPU Arithmetic Benchmark Native FP32 (Float) Whetstone (GFLOPS) 202 [+2%] 198 236 262 Floating-point performance does not change delta either – only 2% faster
CPU Arithmetic Benchmark Native FP64 (Double) Whetstone (GFLOPS) 170 [=] 169 196 223 With FP64 nothing much changes again.
In the legacy integer/floating-point benchmarks ZEN2 is not any faster than ZEN+ despite the change in clocks. Perhaps future microcode updates will help?
BenchCpuMM Native Integer (Int32) Multi-Media (Mpix/s) 1023 [+78%] 574 985 1590 ZEN2 is ~80% faster than ZEN+ despite what we’ve seen before.
BenchCpuMM Native Long (Int64) Multi-Media (Mpix/s) 374 [+2x] 187 414 581 With a 64-bit AVX2 integer vectorised workload, ZEN2 is now 2x faster.
BenchCpuMM Native Quad-Int (Int128) Multi-Media (Mpix/s) 6.56 [+13%] 5.8 6.75 7.56 This is a tough test using Long integers to emulate Int128 without SIMD; here ZEN2 is still 13% faster.
BenchCpuMM Native Float/FP32 Multi-Media (Mpix/s) 100 [+68%] 596 914 1760 In this floating-point AVX/FMA vectorised test, ZEN2 is ~70% faster.
BenchCpuMM Native Double/FP64 Multi-Media (Mpix/s) 618 [+84%] 335 535 533 Switching to FP64 SIMD code, ZEN2 is now ~90% faster than ZEN+
BenchCpuMM Native Quad-Float/FP128 Multi-Media (Mpix/s) 24.22 [+55%] 15.6 23 40.3 In this heavy algorithm using FP64 to mantissa extend FP128, ZEN2 is still 55% faster
With its brand-new 256-bit SIMD units, ZEN2 is anywhere from 55% to 100% faster than ZEN+/ZEN1 a huge upgrade from one generation to the next. For SIMD loads upgrading to ZEN2 gives a huge performance uplift.
BenchCrypt Crypto AES-256 (GB/s) 18 [+12%] 16.1 17.63 23 With AES/HWA support all CPUs are memory bandwidth bound  but ZEN2 manages a 12% improvement.
BenchCrypt Crypto AES-128 (GB/s) 18.76 [+17%] 16.1 17.61 23 What we saw with AES-256 just repeats with AES-128; ZEN2 is now 17% faster.
BenchCrypt Crypto SHA2-256 (GB/s) 20.21 [+9%] 18.6 12 26 With SHA/HWA ZEN2 similarly powers through hashing tests leaving Intel in the dust – and is still ~10% faster than ZEN+
BenchCrypt Crypto SHA1 (GB/s) 20.41 [+6%] 19.3 22.9 38 The less compute-intensive SHA1 does not change things due to acceleration.
BenchCrypt Crypto SHA2-512 (GB/s) 3.77 9 21
ZEN2 with AES/SHA HWA is memory bound like all other CPUs, but it still manages 6-17% better performance than ZEN+ using the same memory. But as ZEN2 is rated for faster memory – using such memory would greatly improve the results.
BenchFinance Black-Scholes float/FP32 (MOPT/s) 257 276 309
BenchFinance Black-Scholes double/FP64 (MOPT/s) 229 [+5%] 219 238 277 Switching to FP64 code, ZEN2 is just 5% faster.
BenchFinance Binomial float/FP32 (kOPT/s) 107 59.9 70.5 Binomial uses thread shared data thus stresses the cache & memory system;
BenchFinance Binomial double/FP64 (kOPT/s) 57.98 [-4%] 60.6 61.6 68 With FP64 code ZEN2 is 4% slower.
BenchFinance Monte-Carlo float/FP32 (kOPT/s) 54.2 56.5 63 Monte-Carlo also uses thread shared data but read-only thus reducing modify pressure on the caches;
BenchFinance Monte-Carlo double/FP64 (kOPT/s) 46.34 [+13%] 41 44.5 50.5 Switching to FP64 nothing much changes, ZEN2 is 13% faster.
Ryzen always did well on non-SIMD floating-point algorithms and here it does not disappoint: ZEN2 does not improve much and is pretty much tied with ZEN+ – thus for non SIMD workloads you might as well stick with the older versions.
BenchScience SGEMM (GFLOPS) float/FP32 263 [-12%] 300 375 413 In this tough vectorised algorithm ZEN2 is strangely slower.
BenchScience DGEMM (GFLOPS) double/FP64 193 [+63%] 119 209 212 With FP64 vectorised code, ZEN2 comes back to be over 60% faster.
BenchScience SFFT (GFLOPS) float/FP32 22.78 [+2.5x] 9 22.33 28.6 FFT is also heavily vectorised but stresses the memory sub-system more; ZEN2 is 2.5x (times) faster.
BenchScience DFFT (GFLOPS) double/FP64 11.16 [+41%] 7.92 11.21 14.6 With FP64 code, ZEN2 is ~40% faster.
BenchScience SNBODY (GFLOPS) float/FP32 612 [+2.2x] 280 557 638 N-Body simulation is vectorised but fewer memory accesses; ZEN2 is over 2x faster.
BenchScience DNBODY (GFLOPS) double/FP64 220 [+2x] 113 171 195 With FP64 precision ZEN2 is almost 2x faster.
With highly vectorised SIMD code ZEN2 improves greatly over ZEN2 sometimes managing to be over 2x faster using the same memory.
CPU Image Processing Blur (3×3) Filter (MPix/s) 2049 [+42%] 1440 2560 4880 In this vectorised integer workload ZEN2 starts over 40% faster than ZEN+.
CPU Image Processing Sharpen (5×5) Filter (MPix/s) 950 [+52%] 627 1000 1920 Same algorithm but more shared data makes ZEN2 over 50% faster.
CPU Image Processing Motion-Blur (7×7) Filter (MPix/s) 495 [+52%] 325 519 1000 Again same algorithm but even more data shared still 50% faster
CPU Image Processing Edge Detection (2*5×5) Sobel Filter (MPix/s) 826 [+67%] 495 827 1500 Different algorithm but still vectorised workload ZEN2 is almost 70% faster.
CPU Image Processing Noise Removal (5×5) Median Filter (MPix/s) 89.68 [+24%] 72.1 78 221 Still vectorised code now ZEN2 drops to just 25% faster.
CPU Image Processing Oil Painting Quantise Filter (MPix/s) 25.05 [+5%] 23.9 42.2 66.7 This test has always been tough for Ryzen so ZEN2 does not improve much.
CPU Image Processing Diffusion Randomise (XorShift) Filter (MPix/s) 1763 [+76%] 1000 4000 4070 With integer workload, Intel CPUs seem to do much better but ZEN2 is still almost 80% faster.
CPU Image Processing Marbling Perlin Noise 2D Filter (MPix/s) 321 [+32%] 243 596 777 In this final test again with integer workload ZEN2 is 32% faster
As we’ve seen before, the new SIMD units are anywhere from 5% (worst-case) to 2x faster than ZEN+/1, a huge performance improvement.
Aggregate Score (Points) 8,200 [+40%] 5,850 7,930 11,810 Across all benchmarks, ZEN2 is ~40% faster than ZEN+.
Aggregating all the various scores, the result was never in doubt: ZEN2 (3700X) is 40% faster than the old ZEN+ (2700X) that itself improved over the original 1700X.

ZEN2’s 256-bit wide SIMD units are a big upgrade and show their power in every SIMD workload; otherwise there is only minor improvement.

SiSoftware Official Ranker Scores

Final Thoughts / Conclusions

Executive Summary: For SIMD workloads you really have to upgrade to Ryzen2; otherwise stick with Ryzen+ unless lower power is preferred. 9/10 overall.

The big change in Ryzen2 are the 256-bit wide SIMD units and all vectorised workloads (Multi-Media, Scientific, Image processing, AI/Machine Learning, etc.) using AVX/FMA will greatly benefit – anything between 50-100% which is a significant increase from just one generation to the next.

But for all other workloads (e.g. Financial, legacy, etc.) there is not much improvement over Ryzen+/1 which were already doing very well against competition.

Naturally it all comes at lower TDP (65W vs 95) which may help with overclocking and also lower noise (from the cooling system) and power consumption (if electricity is expensive or you are running it continuously) thus the performance/W(att) is still greatly improved.

Overall the 3700X does represent a decent improvement over the old 2700X (which is no slouch and was a nice upgrade over 1700X due to better Turbo speeds) and should still be usable in older AM4 300/400-series mainboards with just a BIOS upgrade (without PCIe 4.0).

However, while 2700X (and 1700X/1800X) were top-of-the-line, 3700X is just middle-ground, with the new top CPUs being the 3900X and even the 3950X with twice (2x) more cores and thus potentially huge performance rivaling HEDT Threadripper. The goad-posts have thus moved and thus far higher performance can be yours with just upgrading the CPU. The future is bright…

AMD Ryzen2 3900X Review & Benchmarks – CPU 12-core/24-thread Performance

What is “Ryzen2” ZEN2?

AMD’s Zen2 (“Matisse”) is the “true” 2nd generation ZEN core on 7nm process shrink while the previous ZEN+ (“Pinnacle Ridge”) core was just an optimisation of the original ZEN (“Summit Ridge”) core that while socket compatible it introduces many design improvements over both previous cores. An APU version (with integrated “Navi” graphics) is scheduled to be launched later.

While new chipsets (500 series) will also be introduced and required to support some new features (PCIe 4.0), with an BIOS/firmware update older boards may support them thus allowing upgrades to existing systems adding more cores and thus performance. [Note: older boards will not be enabled for PCIe 4.0 after all]

The list of changes vs. previous ZEN/ZEN+ is extensive thus performance delta is likely to be very different also:

  • Built around “chiplets” of up to 2 CCX (“core complexes”) each of 4C/8T and 8MB L3 cache (7nm)
  • Central I/O hub with memory controller(s) and PCIe 4.0 bridges connected through IF (“Infinity Fabric”) (12nm)
  • Up to 2 chiplets on desktop platform thus up to 2x2x4C (16C/32T 3950X) (same amount as old ThreadRipper 1950X/2950X)
  • 2x larger L3 cache per CCX thus up to 2x2x16MB (64MB) L3 cache (3900X+)
  • 24 PCIe 4.0 lanes (2x higher transfer rate over PCIe 3.0)
  • 2x DDR4 memory controllers up to 4266Mt/s

AMD Ryzen2 3950X chiplets

What’s new in the Ryzen2 core?

Micro-architecturally there are more changes that should improve performance:

  • 256-bit (single-op) SIMD units 2x Fmacs (fixing a major deficiency in ZEN/ZEN+ cores)
  • TLB (2nd level) increased (should help out-of-page access latencies that are somewhat high on ZEN/ZEN+)
  • Memory latencies claim to be reduced through higher-speed memory (note all requests go through IF to Central I/O hub with memory controllers)
  • Load/Store 32bytes/cycle (2x ZEN/ZEN+) to keep up with the 256-bit SIMD units (L1D bandwidth should be 2x)
  • L3 cache is 2x ZEN/ZEN+ but higher latency (cache is exclusive)
  • Infinity Fabric is 512-bit (2x ZEN/ZEN+) and can run 1x or 1/2x vs. DRAM clock (when higher than 3733Mt/s)
  • AMD processors have thankfully not been affected by most of the vulnerabilities bar two (BTI/”Spectre”, SSB/”Spectre v4″) that have now been addressed in hardware.
  • HWM-P (hardware performance state management) transitions latencies reduced (ACPI/CPPCv2)

In this article we test CPU core performance; please see our other articles on:

Hardware Specifications

We are comparing the top-of-the-range Ryzen2 (3900X, 3700X) with previous generation Ryzen+ (2700X) and competing architectures with a view to upgrading to a mid-range high performance design.

CPU Specifications AMD Ryzen 9 3900X (Matisse)
AMD Ryzen 7 3700X (Matisse) AMD Ryzen 7 2700X (Pinnacle Ridge) Intel i9 9900K (Coffeelake-R) Intel i9 7900X (Skylake-X) Comments
Cores (CU) / Threads (SP) 12C / 24T 8C / 16T 8C / 16T 8C / 16T 10C / 20T Matching core-count with CFL (3800X) but 3900X has 50% more cores – more than SKL-X.
Topology 2 chiplets, each 2 CCX, each 3 cores (1 disabled) (12C) 1 chiplet, 2 CCX, each 4 cores (8C) 2 CCX, each 4 cores (8C) Monolithic die Monolithic die AMD uses discrete dies/chiplets unlike Intel
Speed (Min / Max / Turbo) 3.8 / 4.6GHz 3.6 / 4.4GHz 3.7 / 4.2GHz 3.6 / 5GHz 3.3 / 4.3GHz Base clock and turbo are competitive with 3800X having higher base while 3900X higher turbo.
Power (TDP / Turbo) 105 / 135W 65 / 90W 105 / 135W 95 / 135W 140 / 308W TDP remains the same but 3900X may exceed that having more cores.
L1D / L1I Caches 12x 32kB 8-way / 12x 32kB 8-way 8x 32kB 8-way / 8x 32kB 8-way 8x 32kB 8-way / 8x 64kB 4-way 8x 32kB 8-way / 8x 32kB 8-way 10x 32kB 8-way / 10x 32kB 8-way ZEN2 matches L1I with CFL/SKL-X (1/2x ZEN+ but 8-way), L1D is unchanged (also matches Intel)
L2 Caches 12x 512kB (6MB) 8-way 8x 512kB (4MB) 8-way 8x 512kB (4MB) 8-way 8x 256kB (2MB) 16-way 10x 1MB (10MB) 16-way No changes to L2, still 2x CFL. Only SKL-X has its massive 1MB L2 per core which 3900X almost matches!
L3 Caches 2x2x 16MB (64MB) 16-way 2x 16MB (32MB) 16-way 2x 8MB (16MB) 16-way 16MB 16-way 13.75MB 11-way L3 is 2x ZEN/ZEN+ and thus 2x CFL (3800X) with 3900X having a massive 64MB unheard of on the desktop platform! SKL-X can’t match it either.
Mitigations for Vulnerabilities BTI/”Spectre”, SSB/”Spectre v4″ hardware BTI/”Spectre”, SSB/”Spectre v4″ hardware BTI/”Spectre”, SSB/”Spectre v4″ software/firmware RDCL/”Meltdown”, L1TF hardware, BTI/”Spectre”, MDS/”Zombieload”, software/firmware RDCL/”Meltdown” , L1TF, BTI/”Spectre”, MDS/”Zombieload”, all software/firmware Ryzen2 addresses the remaining 2 vulnerabilities while Intel was forced to add MDS to its long list…
Microcode MU-8F7100-11 MU-8F7100-11 MU-8F0802-04 MU-069E0C-9E MU-065504-49 The latest microcodes included in the respective BIOS/Windows have been loaded.
SIMD Units 256-bit AVX/FMA3/AVX2 256-bit AVX/FMA3/AVX2 128bit AVX/FMA3/AVX2 256-bit AVX/FMA3/AVX2 512-bit AVX512 ZEN2 finally matches Intel/CFL but SKL-X’s secret weapon is AVX512 with even consumer CPUs able to do 2x 512-bit FMA ops.

Native Performance

We are testing native arithmetic, SIMD and cryptography performance using the highest performing instruction sets (AVX2, FMA3, AVX, etc.). Ryzen2 supports all modern instruction sets including AVX2, FMA3 and even more like SHA HWA but not AVX-512.

Results Interpretation: Higher values (GOPS, MB/s, etc.) mean better performance.

Environment: Windows 10 x64, latest AMD and Intel drivers. 2MB “large pages” were enabled and in use. Turbo / Boost was enabled on all configurations. All mitigations for vulnerabilities (Meltdown, Spectre, L1TF, MDS, etc.) were enabled as per Windows default where applicable.

Native Benchmarks AMD Ryzen 9 3900X (Matisse)
AMD Ryzen 7 2700X (Pinnacle Ridge)
Intel i9 9900K (Coffeelake-R)
Intel i9 7900X (Skylake-X)
Comments
CPU Arithmetic Benchmark Native Dhrystone Integer (GIPS) 551 [+38%] 334 400 485 Right off Ryzen2 demolishes all CPUs, it is 40% faster than CFL-R!
CPU Arithmetic Benchmark Native Dhrystone Long (GIPS) 556 [+41%] 335 393 485 With a 64-bit integer workload nothing much changes.
CPU Arithmetic Benchmark Native FP32 (Float) Whetstone (GFLOPS) 331 [+40%] 198 236 262 Floating-point performance does not change delta either – still 40% faster!
CPU Arithmetic Benchmark Native FP64 (Double) Whetstone (GFLOPS) 280 [+43%] 169 196 223 With FP64 nothing much changes again.
Ryzen2 starts with an astonishing display, with 3900X demolishing both 9900X and 7900X winning all tests by a large margin 38-43%! It does have 50% more cores (12 vs. 8) but it is not easy to realise gains just by increasing core counts. Intel will need to add far more cores in future CPUs in order to compete!
BenchCpuMM Native Integer (Int32) Multi-Media (Mpix/s) 1449 [+47%] 574 985 1590 Ryzen2 starts off by blowing CFL-R away by 47% and almost matching SKL-X with AVX512!
BenchCpuMM Native Long (Int64) Multi-Media (Mpix/s) 553 [+34%] 187 414 581 With a 64-bit AVX2 integer vectorised workload, Ryzen2 is still 34% faster!
BenchCpuMM Native Quad-Int (Int128) Multi-Media (Mpix/s) 9.52 [+41%] 5.8 6.75 7.56 This is a tough test using Long integers to emulate Int128 without SIMD; here Ryzen2 is again 41% faster!
BenchCpuMM Native Float/FP32 Multi-Media (Mpix/s) 1480 [+62%] 596 914 1760 In this floating-point AVX/FMA vectorised test, Ryzen2 is now over 60% faster than CFL-R and not far off SKL-X!
BenchCpuMM Native Double/FP64 Multi-Media (Mpix/s) 906 [+69%] 335 535 533 Switching to FP64 SIMD code, Ryzen2 is now 70% faster even beating SKL-X!!!
BenchCpuMM Native Quad-Float/FP128 Multi-Media (Mpix/s) 35.23 [+53%] 15.6 23 40.3 In this heavy algorithm using FP64 to mantissa extend FP128, Ryzen2 is still 53% faster!
With its brand-new 256-bit SIMD units, Ryzen2 finally goes toe-to-toe with Intel, soundly beating CFL-R in all benchmarks (+34-69%) sometimes by more than just core count increase (+50%). Only SKL-X with AVX512 manages to be faster (but also with its extra 2 cores). Intel had better release AVX512 for desktop soon but even that will not be enough without increasing core counts to match AMD.
BenchCrypt Crypto AES-256 (GB/s) 15.44 [-12%] 16.1 17.63 23 With AES/HWA support all CPUs are memory bandwidth bound – thus Ryzen2 scores less than Ryzen+ and CFL-R.
BenchCrypt Crypto AES-128 (GB/s) 15.44 [-12%] 16.1 17.61 23 What we saw with AES-256 just repeats with AES-128; Ryzen2 is again slower by 12%.
BenchCrypt Crypto SHA2-256 (GB/s) 29.84 [+2.5x] 18.6 12 26 With SHA/HWA Ryzen2 similarly powers through hashing tests leaving Intel in the dust – 2.5x faster than CFL-R and beating SKL-X with AVX512!
BenchCrypt Crypto SHA1 (GB/s) 19.3 22.9 38
BenchCrypt Crypto SHA2-512 (GB/s) 3.77 9 21
Ryzen2 with AES/SHA HWA is memory bound thus needs faster memory than 3200Mt/s in order to feed all the cores; otherwise due to increased contention for the same bandwidth it may end up slower than Ryzen+ and Intel designs. Here you see the need for HEDT platforms and thus ThreadRipper but at much increased cost.
BenchFinance Black-Scholes float/FP32 (MOPT/s) 257 276 309
BenchFinance Black-Scholes double/FP64 (MOPT/s) 379 [+55%] 219 238 277 Switching to FP64 code, nothing much changes, Ryzen2 55% faster than CFL-R.
BenchFinance Binomial float/FP32 (kOPT/s) 107 59.9 70.5 Binomial uses thread shared data thus stresses the cache & memory system;
BenchFinance Binomial double/FP64 (kOPT/s) 95.73 [+55%] 60.6 61.6 68 With FP64 code Ryzen2 is still 55% faster!
BenchFinance Monte-Carlo float/FP32 (kOPT/s) 54.2 56.5 63 Monte-Carlo also uses thread shared data but read-only thus reducing modify pressure on the caches;
BenchFinance Monte-Carlo double/FP64 (kOPT/s) 76.72 [+72%] 41 44.5 50.5 Switching to FP64 nothing much changes, Ryzen2 is 70% faster than CFL-R and still beating SKL-X.
Ryzen always did well on non-SIMD floating-point algorithms and here it does not disappoint: Ryzen2 is over 50% faster than CFL-R (+55-72%) and soundly beats SKL-X too! As before for financial algorithms there is only one choice and that is Ryzen, be it Ryzen1, Ryzen+ or Ryzen2!
BenchScience SGEMM (GFLOPS) float/FP32 300 375 413 In this tough vectorised algorithm Ryzen2.
BenchScience DGEMM (GFLOPS) double/FP64 212 [+1%] 119 209 212 With FP64 vectorised code, Ryzen2 matches CFL-R and SKL-X.
BenchScience SFFT (GFLOPS) float/FP32 9 22.33 28.6 FFT is also heavily vectorised but stresses the memory sub-system more;
BenchScience DFFT (GFLOPS) double/FP64 12.69 [+13%] 7.92 11.21 14.6 With FP64 code, Ryzen2 is 13% faster than CFL-R.
BenchScience SNBODY (GFLOPS) float/FP32 280 557 638 N-Body simulation is vectorised but fewer memory accesses;
BenchScience DNBODY (GFLOPS) double/FP64 332 [+94%] 113 171 195 With FP64 precision Ryzen2 is almost 2x faster than CFL-R.
With highly vectorised SIMD code Ryzen2 remains competitive but finds some algorithms tougher than others. The new 256-bit SIMD units help but it seems the cores are starved of bandwidth (especially due to SMT) and some workloads may perform better with SMT off.
CPU Image Processing Blur (3×3) Filter (MPix/s) 3056 [+20%] 1440 2560 4880 In this vectorised integer workload Ryzen2 is 20% faster than CFL-R.
CPU Image Processing Sharpen (5×5) Filter (MPix/s) 1499 [+50%] 627 1000 1920 Same algorithm but more shared data makes Ryzen2 50% faster!
CPU Image Processing Motion-Blur (7×7) Filter (MPix/s) 767 [+48%] 325 519 1000 Again same algorithm but even more data shared still 50% faster
CPU Image Processing Edge Detection (2*5×5) Sobel Filter (MPix/s) 1298 [+57%] 495 827 1500 Different algorithm but still vectorised workload Ryzen2 is almost 60% faster.
CPU Image Processing Noise Removal (5×5) Median Filter (MPix/s) 136 [+74%] 72.1 78 221 Still vectorised code now Ryzen2 is 70% faster.
CPU Image Processing Oil Painting Quantise Filter (MPix/s) 38.23 [-9%] 23.9 42.2 66.7 This test has always been tough for Ryzen but Ryzen2 is competitive.
CPU Image Processing Diffusion Randomise (XorShift) Filter (MPix/s) 1384 [-65%] 1000 4000 4070 With integer workload, Intel CPUs seem to do much better.
CPU Image Processing Marbling Perlin Noise 2D Filter (MPix/s) 487 [-18%] 243 596 777 In this final test again with integer workload Ryzen2 is 20% slower.
Thanks to AVX512 SKL-X does win all tests but Ryzen2 beats CFL-R between 20-74% with a few test mixing integer & floating-point SIMD instructions seemingly heavily favouring Intel but nothing to worry about. Overall for image processing Ryzen2 should be your 1st choice.
Aggregate Score (Points) 10,250 [+29%] 5,850 7,930 11,810 Across all benchmarks, Ryzen2 is ~30% faster than CFL-R!
Aggregating all the various scores, the result was never in doubt: Ryzen2 (3900X) is almost 2x faster than Ryzen+ (2700X) and 30% faster than CFL-R, almost catching up HEDT SKL-X.

Ryzen2 (unlike Ryzen1/+) has no trouble with SIMD code due to its widened SIMD units (256-bit) and thus soundly beats the opposition into dust (CFL-R 9900K flagship) sometimes more than just core count increase alone (+50% i.e. 12 cores vs. 8). Sometimes it even beats the AVX512 opposition (SKL-X 7900K) with more cores (10 cores vs. 12).

The only “problematic” algorithms are the memory bound ones where the cores/threads (due to SMT we have 24!) are starved for data and due to contention we see performance lower than less-core devices. While larger caches help (thus the massive 4x 16MB L3 caches) higher clocked memory should be used to match the additional core requirements.

SiSoftware Official Ranker Scores

Final Thoughts / Conclusions

Executive Summary: Ryzen2 is phenomenal and a huge upgrade over Ryzen1/+ that (most) AM4 users can enjoy and Intel has no answer to. 10/10.

Just as original Ryzen forced Intel to increase (double really) core counts to match (from 4 to 6 then 8), Ryzen2 will force Intel to come up with even more (and better) cores in order to compete. 3900X with its 12-cores soundly beats CFL-R 9900K (8-cores) in just about all benchmarks and in some tests goes toe-to-toe with HEDT SKL-X AVX512-enabled (10-cores) except in memory-bound algorithms where the 4 DDR4 memory channels with 2x more bandwidth count. For that you need ThreadRipper!

Ryzen1/+ was already competitive with Intel on integer and floating-point (non-SIMD) workloads but would fare badly on SIMD (AVX/FMA3/AVX2) workloads due to its 128-bit units; Ryzen2 “fixes” this issue, with its 256-bit units matching Intel. Only SKL-X with its 512-bit units (AVX512) is faster and Intel will have to finally include AVX512 for consumer CPUs in order to compete (IceLake?).

For compute-bound workloads, the forthcoming 3950X with its 16-cores/32-threads brings unprecedented performance to the consumer/desktop segment pretty much unheard of just a few years ago when 4-core/8-threads (e.g. 7700K) were all you could hope for – unless paying a lot more for HEDT where 8/10-core CPUs were far far more expensive. Naturally we shall see how the reduced memory bandwidth affects its performance with likely very fast DDR4 memory (4300Mt/s+) required for best performance.

Let’s also remember than Ryzen2 adds hardware mitigation to its remaining 2 vulnerabilities while Intel has been forced to add MDS/”Zombieload” even to its very latest CFL-R that now loses its trump card: hardware RDCL/”Meltdown” fix not to forget the recommendation to disable SMT/Hyperthreading that would mean a sizeable performance drop.

What is astonishing is that TDP has remained similar and with a BIOS/firmware upgrade, owners of older 300-series boards can now upgrade to these CPUs – and likely not even change the cooler unit! Naturally for PCIe4.0 a 500-series board is recommended and 400-series boards do support more features in Ryzen2/+ but let’s remember than on Intel you can only go back/forward 1 generation even though there is pretty much no core difference from Skylake (Gen 6) to Coffeelake-R (Gen 9)!

From top-end (3950X), high-end (3800X) to low-end/APU (3200G) Ryzen2 is such a compelling choice it is hard to recommend anything else… at least at this time…

The new Neural Networks (AI/ML) Benchmarks: RNN Architecture

What is a Recurrent Neural Network (RNN/LSTM)?

A RNN is a type of neural network that is primarily made of of neurons that store their previous states thus are said to ‘have memory’. In effect this allows them to ‘remember’ patterns or sequences.

However, they can still be used as ‘classifiers’ i.e. recognising visual patterns in images and thus can be used in visual recognition software.

What is VGG(net) is why use it now?

VGGNet is the baseline (or benchmark) CNN-type network that while did not win the ILSVRC 2014 competition (won by GoogleNet/Inception) it is still the preferred choice in the community for classification due to its uniform and thus relatively simple architecture.

While it is generally implemented using CNN layers, either directly or combination like ResNet, it can also be implemented using RNN layers which is what we have done here.

We believe this is a good test scenario and thus a relevant benchmark for today’s common systems.

We are considering much complex neurons, like LSTM, for future tests specifically designed for high-end systems as those used in research and academia.

What is the MNIST dataset and why use it now?

The MNIST database (https://en.wikipedia.org/wiki/MNIST_database) is a decently sized dataset of handwritten digits used for training and testing image processing systems like neural networks. It contains 60K training and 10K testing images of 28×28 pixel anti-aliased gray levels. The number of classes is only 10 (digits ‘0’ to ‘9’).

While they are only 28×28 and not colour, they can be up-scaled to any size by common up-scaling algorithms to test neural networks with little source data.

Today (2019) the digits would be captured in much higher resolution similar to the standard input resolution of the image processing networks of today (between 200×200 and 300×300 pixels).

As Sandra is designed to be small and easily downloadable, it is not possible to include gigabytes (GB) of data for either inference or training. Even the low-resolution (32x32x3) ILSVRC is 3GB thus unusable for our purpose.

What is Sandra’s RNN network architecture and why was it designed this way?

Due to the low complexity of the data and in order to maintain good performance even on low-end hardware, a standard RNN was chosen as the architecture. The features are:

  • Input is 224x224x1 as MNIST images are grey-scale only (up-scaled from 28×28)
  • Output is 10 as there are only 10 classes
  • 4 layer network, 1 RNN, 3 fully connected layers

What are the implementation details of the network?

The CPU version of the neural network supports all common instruction sets and precision and will be continuously updated as the industry moves forward.

  • Both inference/forward and train/back-propagation tested and supported.
  • Precision: single and double floating-point supported with future half/FP16.
  • SIMD Instruction Sets: CPU, SSE2, SSE4.x, AVX, AVX2/FMA and AVX512 with future VNNI.
  • Threads/Cores: Up to the maximum operating system 384 threads in 64-thread groups are supported with hard affinity as all other benchmarks.
  • NUMA: NUMA is supported up to 16 nodes with data allocated to the closest node.

What kind of BTT (Back-propagation Through Time) is used?

Unfortunately as we only know the output (digit) at the end of the sequence (i.e. once all pixels have been presented) we cannot calculate intermediate errors in order to use TBTT (Truncated BTT) which relies on known output at intermediate sequence time-steps.

What kind of detection rate and error does Sandra’s implementation achieve?

Naturally due to the low source resolution, a much shallower/simpler network would have sufficed. However due to up-scaling and the relatively large number of training images there is no danger of over-fitting.

It achieves a % detection rate (over the 10K testing images) after just 1 epoch (Epoch 0) and % after 30 epochs.

Training (30 epochs) took just X* hours on an i9-7900X (10C/20T) using AVX512/single-precision.

Does Sandra fully infer or train the full image set when benchmarking?

As with all other Sandra benchmarks the tests are limited to 30 seconds (in order to complete reasonably quickly) – within this time as many images at random from the data-sets (60K train, 10K test) will be processed.

The new Neural Networks (AI/ML) Benchmarks: CNN Architecture

What is a Convolution Neural Network (CNN/ConvNet)?

A CNN is a type of neural network that is primarily made of of neuron layers connected in such a way that they perform convolution over the previous layers: in effect they are filters over the input – the same way a blur/sharpen/edge/etc filter would be applied over a picture.

They are used as ‘classifiers’ i.e. recognising visual patterns in images and thus are used in visual recognition software.

What is VGG(net) is why use its architecture now?

VGGNet is the baseline (or benchmark) CNN-type network that while did not win the ILSVRC 2014 competition (won by GoogleNet/Inception) it is still the preferred choice in the community for classification due to its uniform and thus relatively simple architecture.

Thus while today (2019) there are far deeper and more complex neural networks, as Sandra is intended to run on common systems we had to choose the most common but relatively simple network.

We believe this is a good test scenario and thus a relevant benchmark for today’s common systems.

We are considering much deeper networks, like ResNet, for future tests specifically designed for high-end systems as those used research and academia.

Why not use Tensorflow, Caffee, etc. as back-end?

As with all Sandra benchmarks we develop our own code which is optimised in the conjunction with the community which includes hardware makers. This allows us to control all the benchmark stack adding new features and support as required which we would not be able to do when using a back-end.

Using a specific vendor’s libraries (e.g. cuDNN, MKL, etc.) would lock-us into a specific platform while we provide implementation for all platforms including all CPU SIMD instruction sets (SSE2, SSE4, AVX, AVX2/FMA, AVX512) and major GP (GPGPU) run-times (CUDA, OpenCL, DirectX 11/12 Compute and future Vulkan*).

What is the MNIST dataset and why use it now?

The MNIST database (https://en.wikipedia.org/wiki/MNIST_database) is a decently sized dataset of handwritten digits used for training and testing image processing systems like neural networks. It contains 60k (thousand) training and 10k testing images of 28×28 pixel anti-aliased gray levels. The number of classes is only 10 (digits ‘0’ to ‘9’).

While they are only 28×28 and not colour (1 channel), they can be up-scaled to any size by common up-scaling algorithms to test neural networks with little source data. Here we up-scale them 8x to 224x224x1.

Today (2018) the digits would be captured in much higher resolution similar to the standard input resolution of the image processing networks of today (between 200×200 and 300×300 pixels).

As Sandra is designed to be small and easily downloadable, it is not possible to include gigabytes (GB) of data for either inference or training. Even the low-resolution ImageNet ILSVRC is 3GB thus unusable for our purpose.

What are the CIFAR datasets and why use them now?

The CIFAR datasets (https://www.cs.toronto.edu/~kriz/cifar.html) are also decently sized datasets of objects used for training and testing image processing systems like neural networks. They both consists of 50k (thousand) training and 10k testing images of 32x32x3 pixel colour images with CIFAR-10 having 10 classes and CIFAR-100 having 100 classes.

Unlike MNIST the pictures are colour (3 channels RGB) and can also be up-scaled to any size by common up-scaling algorithms to test neural networks with little source data. Here we up-scale them 7x to 224x224x3.

Again, just as with MNIST this allows us to include more datasets while processing them in high resolution similar to modern neural networks without including a large dataset like ImageNet ILSVRC dataset.

What are ImageNet ILSVRC datasets and why *not* use them?

The ImageNet (ImageNet Large Scale Visual Recognition Challenge) datasets (http://www.image-net.org/challenges/LSVRC/) are used in the yearly challenge for researchers in object detection, image classification at large scale. They are used to measure progress in computer vision in the World today.

The yearly challenge/competition has thus yielded many recent advancements in the field with winners (and in some cases runner-ups) providing the classical neural networks of today: AlexNet, VGG, ResNet, Inception, etc.

Naturally the task is non-trivial and requires cutting-edge complex neural networks that generally require similarly high-end hardware that is not the domain of mass-market. While old(er) neural networks like AlexNet, VGG or ResNet can today (2018) work on consumer hardware – they are usually deployed in inference/classification mode. Training them (from scratch) would still require significant processing power and time which does not make sense for our benchmark.

Due to the nature of our software (mass-market, small, fast) the size of the datasets (about 3GB for 32x32x3 1.2 million training images) makes them unsuitable to be included either as standard or downloadable. As we aready use low-resolution datasets, it would not make sense to include another – and the high resolution versions (e.g. 256x256x3) are far larger (about 137GB train, 6.3GB test).

Another issue is the licensing: they are licensed for research which Sandra as a commercial product – even though we provide the benchmarks free of charge – would likely not qualify.

What is Sandra’s CNN network architecture and why was it designed this way?

Due to the low complexity of the data and in order to maintain good performance even on low-end hardware, VGG-16 was chosen as the architecture. The features are:

  • For MNIST dataset
    • Input is 224x224x1 as MNIST images are grey-scale (upscaled from 28×28)
    • Output is 10* as there are only 10 classes
    • 8 convolution (3×3 step 1), 5 pooling (2×2 step 2), 3 full-connect layers
  • Network/Engine features
    • Layers: Fully Connected/Dense, Convolution, Max Pooling, Recurrent, Dropout.
    • Activation: ReLU, Leaky ReLU, Smooth ReLU, Sigmoid, TanH. Activation functions are fused to the layers for reduced memory size/bandwidth footprint.
    • Back-propagation Optimiser: 2nd order Hessian.
    • Alignment: For performance, some layer sizes may be increased (e.g. output) to match SIMD alignment; the performance due to SIMD is higher than the overhead due to more un-needed neurons.
    • SIMD Float Width: Up to 64 single-precision pixels per cycle when using AVX512.
    • SIMD Half Width: Up to 128 half-precision pixels per cycle when using AVX512/BFloat16*.
    • SIMD Int8 Width: Up to 256 int8 pixels per cycle when using AVX512/VNNI*.

What are the implementation details of the network?

The CPU version of the neural network supports all common instruction sets and precision and will be continuously updated as the industry moves forward.

  • Both inference/forward and train/back-propagation tested and supported.
  • Processor:
    • Precision: single/FP32 and double/FP64 supported.
    • SIMD Instruction Sets: FPU, SSE2, SSE4.x, AVX, AVX2/FMA, AVX512 with future VNNI*.
    • Threads/Cores: Up to the maximum operating system 384 threads in 64-thread groups are supported with hard affinity as all other benchmarks.
    • Atomic Updates: TSX/RTE used where supported otherwise 128/64/32-bit interlock/update.
    • NUMA: NUMA is supported up to 16 nodes with data allocated to the closest node.
    • Large Pages: Large (2/4MB) pages used where supported and enabled.
  • GP (GPGPU):
    • Precision: single/FP32 and half/FP16 supported.
    • Run-Times: CUDA 10+, OpenCL 1.2+, DirectX 11/12 Compute.
    • Multi-GPU: Up to 8 devices are supported including CPU pseudo-device.

How is the data stored/processed?

We use the CHW format for simple SIMD implementation and performance load/store.

What activation function do you use?

We use the Sigmoid activation function with a fast (but naturally somewhat low-precision) SIMD tanh/exp implementation; while many modern networks (and VGG itself) use ReLU (for speed reasons) we’ve found the Sigmoid to work “better” for us without appreciable performance impact. By better we mean fast convergence and no need for batch normalisation.

What kind of detection rate and error does Sandra’s implementation achieve?

Naturally due to the low source resolution, a much shallower/simpler network would have sufficed. However due to upscaling and the relatively large number of training images there is no danger of overfitting.

It achieves a 95.3% detection rate (over the 10k testing images) after just 1 epoch (Epoch 0) and 99.82% after 30 epochs.

Training (30 epochs) took just 7* hours on an i9-7900X (10C/20T) using AVX512/single-precision.

Does Sandra fully infer or train the full image set when benchmarking?

As with all other Sandra benchmarks the tests are limited to 30 seconds (in order to complete resonably quickly) – within this time as many images at random from the datasets (60k train, 10k test) will be processed.

SiSoftware Sandra Titanium (2018) SP4/a/c Update: Retpoline and hardware support

Note: Updated 2019/June with information regarding MDS as well as change of recent CFL-R microcode vulnerability reporting.

We are pleased to release SP4/a/c (version 28.69) update for Sandra Titanium (2018) with the following updates:

Sandra Titanium (2018) Press Release

  • Reporting of Operating System (Windows) speculation control settings for the recently discovered vulnerabilities:
    • Kernel Retpoline mitigation status (for RDCL) in recent Windows 10 / Server 2019 updates
    • Kernel Address Table Import Optimisation (“KATI”) status (as above)
    • L1TFL1 data terminal fault mitigation status
    • MDSMicroarchitectural Data Sampling/”ZombieLoad” mitigation status
  • Hardware Support:
    • AMD Ryzen2 (Matisse), Stoney Ridge support
    • Intel CometLake (CML), CannonLake (CNL), IceLake (ICL) support (based on public information)
  • CPU Benchmarks:
    • Image Processing: SIMD code improvement (SSE2/SSE4/AVX/AVX2-FMA/AVX512)
    • Multi-Media: Lock-up on NUMA systems (e.g. AMD ThreadRipper) thanks to Rob @ TechGage.
  • Memory/Cache Benchmarks
    • Return memory controller firmware version to Ranker
  • GPGPU Benchmarks:
    • CUDA SDK 10.1
    • OpenCL: Processing (Fractals/Mandelbrot) variable vector width based on reported FP16/32/64 optimal SIMD width.
  • Ranker, Price & Information Engines
    • HTTPS (encryption) support for all engines as well as the main website

What is Retpoline?

It is a mitigation against ‘Spectre‘ 2 variant (BTI – Branch Target Injection) that affects just about all CPUs (not just Intel but AMD, ARM, etc.). While ‘Spectre’ does not have the same overall performance impact degradation as ‘Meltdown‘ (RDCL – Rogue Data Cache Load) it can have a sizeable impact on some processors and workloads. At this time no CPUs contain hardware mitigation for Spectre without performance impact.

Retpoline (Return Trampoline) is a faster way to mitigate against it without restricting branch speculation in kernel mode (using IBRS/IBPB) and has recently been added to Linux and now Windows version 1809 builds with KB4482887. Note that it still needs to be enabled in registry via the Mitigation Features Override flags as by default it is not enabled.

What CPUs can Retpoline be used on?

Unfortunately Retpoline is only safe to use on some CPUs: AMD CPUs (though does not engage on Ryzen, see below), Intel Broadwell or older (v5 and earlier) – thus not Skylake (v6 or later).

Windows speculation control settings reporting:

Intel Haswell (Core v4), Broadwell (v5) – Retpoline enabled, KATI enabled
Kernel Retpoline Speculation Control – Enabled

Kernel Address Table Import Optimisation – Enabled

(Note RDCL mitigations KVA, L1TF are also enabled as required)

Intel Skylake (Core v6), Kabylake (v7), Skylake/Kabylake-X (v6x) – no Retpoline, KATI can be enabled
Kernel Retpoline Speculation Control – no

Kernel Address Table Import Optimisation – no/yes (can be enabled)

(Note RDCL mitigations KVA, L1TF are enabled as required)

Intel Coffeelake-R (Core v8r), Whiskeylake/AmberLake (Core v8r), CometLake* – no Retpoline, KATI not enabled
Kernel Retpoline Speculation Control – no

Kernel Address Table Import Optimisation – Enabled

Note 2019/June: Latest microcode (AEh) with MDS vulnerability support cause Windows to report KVA/L1TF mitigations as required despite CPU claiming to not be vulnerable to RDCL.

Intel Atom Braswell (Atom v5), GeminiLake/ApolloLake (Atom v6) – no Retpoline but KATI enabled
Kernel Retpoline Speculation Control – no

Kernel Address Table Import Optimisation – Enabled

(Note RDCL mitigations KVA, L1TF are enabled as required)

AMD Ryzen (Threadripper) 1, 2 – no Retpoline, no KATI
Kernel Retpoline Speculation Control – no (should be usable?)

Kernel Address Table Import Optimisation – no (should be usable)

(Note CPU does not require RDCL mitigation thus no KVA, L1TF required)

From our somewhat limited testing above it seems that:

  • Intel Haswell/Broadwell (Core v4/v5) and perhaps earlier (Ivy Bridge/Sandy Bridge Core v3/v2) users are in luck, Retpoline is enabled and should improve performance; unfortunately RDCL (“Meltdown” mitigation) remains.
  • Intel Coffeelake-R (Core v8r refresh), Whiskylake ULV (v8r) users do benefit a bit more for their investment – while Retpoline is not enabled, KATI is enabled and should help. Not requiring KVA is the biggest gain of CFL-R. 2019/June: latest microcode (AEh) causes Windows to require KVA/L1TF thus negating any benefit CFL-R had over original CFL/KBL/SKL.
  • Intel Skylake (Core v6), Kabylake (v7) and Coffeelake (v8) are not able to benefit from Retpoline but KATI can work on some systems (driver dependent). However, on our Skylake ULV, Skylake-X test systems KATI could not be enabled. We are investigating further.
  • Intel Atom (v4/v5+) users should be able to use Retpoline but it seems it cannot be enabled currently. KATI is enabled.
  • AMD Ryzen (Threadripper) 1/2 users should also be able to use Retpoline but it seems it cannot be enabled currently. While RDCL is not required, mitigations for Spectre v2 are required and should be enabled. We are investigating further.

Reviews using Sandra 2018 SP4:

Update & Download

Commercial version customers can download the free updates from their software distributor; Lite users please download from your favourite download site.

Download Sandra Lite

AMD Ryzen 2 Mobile (2500U) Vega 8 GP(GPU) Performance

What is “Ryzen2” ZEN+ Mobile?

It is the long-awaited Ryzen2 APU mobile “Bristol Ridge” version of the desktop Ryzen 2 with integrated Vega graphics (the latest GPU architecture from AMD) for mobile devices. While on desktop we had the original Ryzen1/ThreadRipper – there was no (at least released) APU version or a mobile version – leaving only the much older designs that were never competitive against Intel’s ULV and H APUs.

After the very successful launch of the original “Ryzen1”, AMD has been hard at work optimising and improving the design in order to hit TDP (15-35W) range for mobile devices. It has also added the brand-new Vega graphics cores to the APU that have been incredibly performant in the desktop space. Note that mobile versions have a single CCX (compute unit) thus do not require operating system kernel patches for best thread scheduling/power optimisation.

Here’s what AMD says it has done for Ryzen2 mobile:

  • Process technology optimisations (12nm vs 14nm) – lower power but higher frequencies
  • Radeon RX Vega graphics core (DirectX 12.1)
  • Optimised boost (aka Turbo) algorithm – sharing between CPU & GPU cores

In this article we test GP(GPU) integrated graphics performance; please see our other articles on:

Hardware Specifications

We are comparing the graphics units of Ryzen2 mobile with competitive APUs with integrated graphics  to determine whether they are good enough for modest use, especially for compute (GPGPU) use supporting the CPU.

GPGPU Specifications AMD Radeon RX Vega 8 (2500U)
Intel UHD 630 (7200U)
Intel HD Iris 520 (6500U)
Intel HD Iris 540 (6550U)
Comments
Arch Chipset GCN1.5 GT2 / EV9.5 GT2 / EV9 GT3 / EV9 All graphics cores are minor revisions of previous cores with extra functionality.
Cores (CU) / Threads (SP) 8 / 512 24 / 192 24 / 192 48 / 384 Vega has the most SPs though only a few but powerful CUs
ROPs / TMUs 8 / 32 8 / 16 8 / 16 16 / 24 Vega has less ROPs than GT3 but more TMUs.
Speed (Min-Turbo) 300-1100 300-1000 300-1000 300-950 Turbo boost puts Vega in top position power permitting.
Power (TDP) 25-35W 15-25W 15-25W 15-25W TDP is about the same for all though both Ryzen2 and CFL-U have somewhat higher TDP (25W).
Constant Memory 2.7GB 1.6GB 1.6GB 3.2GB There is no dedicated constant memory thus a large chunk is available to use (GB) unlike a dedicated video card with very fast but small (kB).
Shared (Local) Memory 32kB 64kB 64kB 64kB Intel has 2x larger shared/local memory but slow (likely non dedicated) unlike Vega.
Global Memory 2.7 / 3GB 1.6 / 3.2GB 1.6 / 3.2GB 3.2 / 6.4GB About 50% of main memory can be used as global memory – thus pretty large workloads can be run.
Memory System 128-bit DDR4 2400Mt/s 128-bit DDR3L 1866Mt/s 128-bit DDR3L 1866Mt/s 128-bit DDR4 2133MT/s Ryzen2’s memory controller is rated for faster data rates thus should be able to use faster (laptop) memory.
Memory Bandwidth (GB/s)
36 30 30 33 The high data rate of DDR4 can result in higher bandwidth useful for the GPU cores.
L2 Cache ? 512kB 512kB 1MB L2 is comparable to Intel units.
FP64/double ratio Yes, 1/16x Yes, 1/8x Yes, 1/8 Yes, 1/8x FP64 is supported and at good ratio but lower than Intel’s.
FP16/half ratio
Yes, 2x Yes, 2x Yes, 2x Yes, 2x FP16 is also now supported at twice the rate – again unlike gimped dedicated cards.

Processing Performance

We are testing both OpenCL performance using the latest SDK / libraries / drivers from both AMD and competition.

Results Interpretation: Higher values (GOPS, MB/s, etc.) mean better performance.

Environment: Windows 10 x64, latest Intel drivers, OpenCL 2.x. Turbo / Boost was enabled on all configurations.

Processing Benchmarks Intel UHD 630 (7200U) Intel HD Iris 520 (6500U) Intel HD Iris 540 (6550U) AMD Radeon RX Vega 8 (2500U) Comments
GPGPU Arithmetic Benchmark Mandel FP16/Half (Mpix/s) 831 927 1630 2000 [+23%] Thanks to FP16 support we see double the performance over FP32 but Vega is only 23% faster than GT3.
GPGPU Arithmetic Benchmark Mandel FP32/Single (Mpix/s) 476 478 865 1350 [+56%] Vega rules FP32 and is over 50% faster than GT3.
GPGPU Arithmetic Benchmark Mandel FP64/Double (Mpix/s) 113 122 209 111 [-47%] FP64 lower rate makes Vega 1/2 the speed of GT3 and only matching GT2 units.
GPGPU Arithmetic Benchmark Mandel FP128/Quad (Mpix/s) 5.71 6.29 10.78 7.11 [-34%] Emulated FP128 precision depends entirely on FP64 performance thus not a lot changes.
Vega is over 50% faster than Intel’s top-end Iris/GT3 graphics but only in FP32 precision – while it gains from FP16 Intel scales better reducing the lead to just 25% or so. In FP64 precision though it’s relatively low 1/16x ratio means it only ties with GT2 low-end-models while GT3 is 2x (twice) as fast. Pity.
GPGPU Crypto Benchmark Crypto AES-256 (GB/s) 0.858 0.87 1.23 2.58 [+2.1x] No wonder AMD is crypto-king: Vega is over 2x faster than even GT3.
GPGPU Crypto Benchmark Crypto AES-128 (GB/s) 1 1.08 1.52 3.3 [+2.17x] Nothing changes here, Vega is over 2.2x faster.
GPGPU Crypto Benchmark Crypto SHA2-256 (GB/s) 2.72 3 4.7 14.29 [+3x] In this heavy integer workload, Vega is now 3x faster no wonder it’s used for crypto mining.
GPGPU Crypto Benchmark Crypto SHA1 (GB/s) 6 6.64 11.59 18.77 [+62%] SHA1 is less compute intensive allowing Intel to catch up but Vega is still over 60% faster.
GPGPU Crypto Benchmark Crypto SHA2-512 (GB/s) 1.019 1.08 1.86 3.36 [+81%] With 64-bit integer workload, Vega does better and is 80% (almost 2x) faster than GT3.
Nobody will be using integrated graphics for crypto-mining any time soon, but if you needed to (perhaps using encrypted containers, VMs, etc.) then Vega is your choice – even GT3 is left in the dust despite big improvement over low-end GT2. Intel would need at least 2x more cores to be competitive here.
GPGPU Finance Benchmark Black-Scholes half/FP16 (MOPT/s) 1000 1140 1470 1720 [+17%] If 16-bit precision is sufficient for financial work, Vega is 20% faster than GT3.
GPGPU Finance Benchmark Black-Scholes float/FP32 (MOPT/s) 694 697 794 829 [+4%] In this relatively simple FP32 financial workload Vega is just 4% faster than GT3.
GPGPU Finance Benchmark Black-Scholes double/FP64 (MOPT/s) 142 154 281 185 [-33%] Switching to FP64 precision, Vega is 33% slower than GT3.
GPGPU Finance Benchmark Binomial half/FP16 (kOPT/s) 86 95 155 270 [+74%] Switching to 16-bit precision allows Vega to gain over GT3 and is almost 2x faster.
GPGPU Finance Benchmark Binomial float/FP32 (kOPT/s) 92 93 153 254 [+66%] Binomial uses thread shared data thus stresses the internal memory sub-system, and here Vega shows its power – it is 66% faster than GT3.
GPGPU Finance Benchmark Binomial double/FP64 (kOPT/s) 18 18.86 32 15.67 [-51%] With FP64 precision Vega loses again vs. GT3 at 1/2 the speed and just matches GT2 units.
GPGPU Finance Benchmark Monte-Carlo half/FP16 (kOPT/s) 211 236 395 584 [+48%] With 16-bit precision, Vega dominates again and is almost 50% faster than GT3.
GPGPU Finance Benchmark Monte-Carlo float/FP32 (kOPT/s) 223 236 412 362 [-12%] Monte-Carlo also uses thread shared data but read-only thus reducing modify pressure – but Vega somehow loses against GT3.
GPGPU Finance Benchmark Monte-Carlo double/FP64 (kOPT/s) 29.5 33.36 58.7 47.13 [-20%] Switching to FP64 precision as expected Vega is slower.
Financial algorithms perform well on Vega – at least in FP16 & FP32 precision but FP64 is too “gimped” (1/16x FP32 rate) and thus loses against GT3 despite more powerful cores.
GPGPU Science Benchmark HGEMM (GFLOPS) half/FP16 127 140 236 884 [+3.75x] With 16-bit precision Vega runs away with GEMM and is almost 4x faster than GT3.
GPGPU Science Benchmark SGEMM (GFLOPS) float/FP32 105 107 175 214 [+79%] GEMM makes heavy use of shared/local memory which is likely why Vega is 80% faster than GT3.
GPGPU Science Benchmark DGEMM (GFLOPS) double/FP64 38.8 41.69 70 62.6 [-11%] As expected, due to gimped FP64 rate Vega falls behind GT3 but only by just 11%.
GPGPU Science Benchmark HFFT (GFLOPS) half/FP16 34.2 34.7 45.85 61.34 [+34%] 16-bit precision helps reduce memory bandwidth pressure thus Vega is 34% faster.
GPGPU Science Benchmark SFFT (GFLOPS) float/FP32 20.9 21.45 29.69 31.48 [+6%] FFT is memory access bound but Vega does well to beat GT3.
GPGPU Science Benchmark DFFT (GFLOPS) double/FP64 4.3 5.4 6.07 14.19 [+2.34x] Despite the FP64 rate, Vega manages its memory accesses better beating GT3 by over 2x (two times).
GPGPU Science Benchmark HNBODY (GFLOPS) half/FP16 270 284 449 623 [+39%] 16-bit precision still benefits N-Body and here Vega is 40% faster than GT3.
GPGPU Science Benchmark SNBODY (GFLOPS) float/FP32 162 181 291 537 [+85%] Back to FP32 and Vega has a pretty large 85% lead – almost 2x GT3.
GPGPU Science Benchmark DNBODY (GFLOPS) double/FP64 22.73 26.1 43.34 44 [+2%] With FP64 precision, Vega and GT3 are pretty much tied.
Vega performs well on compute heavy scientific algorithms (making heavy use of shared/local memory) and also benefits from half/FP16 to reduce memory bandwidth pressure, but FP64 rate comes back to haunt it where it loses against Intel’s GT3. Pity.
GPGPU Image Processing Blur (3×3) Filter half/FP16 (MPix/s) 888 937 1390 2273 [+64%] With 16-bit precision Vega doubles its lead to 64% over GT3 despite its gain over FP32.
GPGPU Image Processing Blur (3×3) Filter single/FP32 (MPix/s) 461 491 613 781 [+27%] In this 3×3 convolution algorithm, Vega does well but only 30% faster than GT3.
GPGPU Image Processing Sharpen (5×5) Filter half/FP16 (MPix/s) 279 302 409 582 [+42%] Again a huge gain by using FP16, over 40% faster than GT3.
GPGPU Image Processing Sharpen (5×5) Filter single/FP32 (MPix/s) 100 107 144 157 [+9%] Same algorithm but more shared data reduces the gap to 9%.
GPGPU Image Processing Motion Blur (7×7) Filter half/FP16 (MPix/s) 254 272 396 619 [+56%] Large gain again by switching to FP16 with 3x performance over FP32.
GPGPU Image Processing Motion Blur (7×7) Filter single/FP32 (MPix/s) 103 111 156 161 [+3%] With even more shared data the gap falls to just 3%.
GPGPU Image Processing Edge Detection (2*5×5) Sobel Filter half/FP16 (MPix/s) 259 281 363 595 [+64%] Another huge gain and over 3x improvement over FP32.
GPGPU Image Processing Edge Detection (2*5×5) Sobel Filter single/FP32 (MPix/s) 99 106 145 155 [+7%] Still convolution but with 2 filters – the gap is similar to 5×5 – Vega is 7% faster.
GPGPU Image Processing Noise Removal (5×5) Median Filter half/FP16 (MPix/s) 7.39 9.4 8.56 7.688 [-18%] Big gain but not enough to beat GT3 here.
GPGPU Image Processing Noise Removal (5×5) Median Filter single/FP32 (MPix/s) 7 7.57 7.08 4 [-47%] Vega does not like this algorithm (lots of branching causing divergence) and is 1/2 GT3 speed.
GPGPU Image Processing Oil Painting Quantise Filter half/FP16 (MPix/s) 8.55 9.32 9.22 <BSOD> This test would cause BSOD; we are investigating.
GPGPU Image Processing Oil Painting Quantise Filter single/FP32 (MPix/s) 8 8.65 6.77 2.59 [-70%] Vega does not like this algorithms either (complex branching) and neither does GT3.
GPGPU Image Processing Diffusion Randomise (XorShift) Filter half/FP16 (MPix/s) 941 967 1580 2091 [+32%] In order to prevent artifacts most of this test runs in FP32 thus not much gain here.
GPGPU Image Processing Diffusion Randomise (XorShift) Filter single/FP32 (MPix/s) 878 952 1550 2100 [+35%] This algorithm is 64-bit integer heavy allowing Vega 35% better performance over GT3.
GPGPU Image Processing Marbling Perlin Noise 2D Filter half/FP16 (MPix/s) 341 390 343 1046 [+2.5x] Switching to FP16 makes a huge difference to Vega which is over 2x faster.
GPGPU Image Processing Marbling Perlin Noise 2D Filter single/FP32 (MPix/s) 384 425 652 608 [-7%] One of the most complex and largest filters, Vega is a bit slower than GT3 by 7%.
For image processing Vega generally performs well in FP32 beating GT3 hands down; but there are a few algorithms that may need to be optimised for it that don’t perform as well as expected. Switching to FP16 though doubles/triples scores – thus Vega may be starved of memory.

Memory Performance

We are testing both OpenCL performance using the latest SDK / libraries / drivers from both AMD and competition.

Results Interpretation: Higher values (MB/s, etc.) mean better performance. Lower time values (ns, etc.) mean better performance.

Environment: Windows 10 x64, latest Intel drivers, OpenCL 2.x. Turbo / Boost was enabled on all configurations.

Memory Benchmarks Intel UHD 630 (7200U) Intel HD Iris 520 (6500U) Intel HD Iris 540 (6550U) AMD Radeon RX Vega 8 (2500U) Comments
GPGPU Memory Bandwidth Internal Memory Bandwidth (GB/s) 12.17 21.2 24 27.32 [+14%] With higher speed DDR4 memory, Vega has 14% more bandwidth.
GPGPU Memory Bandwidth Upload Bandwidth (GB/s) 6 10.4 11.7 4.74 [-60%] The GPU<>CPU link seems a bit slow here at 1/2 bandwidth of Intel.
GPGPU Memory Bandwidth Download Bandwidth (GB/s) 6 10.5 11.75 5 [-57%] Download bandwidth shows a similar issue, 1/2 bandwidth expected.
All designs have to rely on the shared memory controller and Vega performs as expected with good internal bandwidth due to higher speed DDR4 memory. But – transfer up/down speeds are disappointing possibly due to the driver as “zero-copy” mode should be engaged and working on such transfers (APU mode).
GPGPU Memory Latency Global (In-Page Random Access) Latency (ns) 246 244 288 412 [+49%] Similarly with CPU data latencies, global “in-page/random” (aka “TLB hit”) latencies are a bit high though not by a huge amount.
GPGPU Memory Latency Global (Full Range Random Access) Latency (ns) 365 372 436 519 [+19%] Due to faster memory clock but increased timings “full/random” latencies appear a bit higher.
GPGPU Memory Latency Global (Sequential Access) Latency (ns) 156 158 213 201 [-6%] Sequential access latencies are less than competition by 6%.
GPGPU Memory Latency Constant Memory (In-Page Random Access) Latency (ns) 245 243 252 411 [+63%] None have dedicated constant memory thus we see a similar picture to global memory: somewhat high latencies.
GPGPU Memory Latency Shared Memory (In-Page Random Access) Latency (ns) 82 84 100 22.5 [1/5x] Vega has dedicated shared/local memory and it shows – it’s about 5x faster than Intel’s designs.
GPGPU Memory Latency Texture (In-Page Random Access) Latency (ns) 1152 1157 1500 278 [1/5x] Texture access is also very fast on Vega, with latencies 5x lower (aka 1/5) than Intel’s designs.
GPGPU Memory Latency Texture (Full Range Random Access) Latency (ns) 1178 1162 1533 418 [1/3x] Even full/random accesses are fast, 3x (three times) faster than Intel’s.
GPGPU Memory Latency Texture (Sequential Access) Latency (ns) 1077 1081 1324 122 [1/10x] With sequential access we see a crazy 10x lower latency as if AMD uses prefetchers and Intel does not.
As we’ve seen in Ryzen 2’s data latency tests – “in-page/random” latencies are higher than competition but the rest are comparative, with sequential (prefetched) latencies especially small. But dedicated shared/local memory is far faster (5x) and texture accesses are also very fast (3-5x) which should greatly help algorithms making use of them.
Plotting the global (or constant) memory latencies together we see that the “in-page/random” access latencies should perhaps peak somewhat lower but still nothing close to what we’ve seen in the (CPU) data memory latencies article. It is not very clear (unlike the texture latencies graph) where the caches are located.
The texture latencies graph is far clearer where we can see each level’s caches; unlike the global (or constant) latencies we see “in-page/random” latency peak and hold at a somewhat lower level (4MB).

SiSoftware Official Ranker Scores

Final Thoughts / Conclusions

Vega mobile, as its desktop big siblings, is undoubtedly powerful and a good upgrade from the older integrated GPU cores; it also supports modern features like half/FP16 compute (which needs vectorisation what the driver reports as “optimised width”) and relishes complex algorithms making use of shared/local memory which is efficient. However Intel’s GT3 EV9.x can get close to it in some workloads and due to better FP64 ratio (1/8x vs 1/16x) even beat it in most FP64 precision tests which is somewhat disappointing.

Luckily for AMD, GT3 variant is very rare and thus Vega has an easy job defeating GT2 in just about all tests; but it shows that should Intel “get serious” and continue to improve integrated graphics (and CPUs) like they used to do before Skylake (SKL/KBL) – AMD might have more serious competition on its hands.

Note that until recently (2019) Ryzen2 mobile APUs were not supported by AMD’s main drivers (“Adrenalin”) and had to rely on pretty old OEM (HP, etc.) drivers that were somewhat problematic especially with Windows 10 changing every 6 months while the drivers were almost 1 year old. Thankfully this has now changed and users (and us) can benefit from updated, stable and performant drivers.

In any case if you want a laptop/ultraportable with just an APU and no dedicated graphics, then Vega is pretty much your only choice which means a Ryzen2 system. That pretty much means it is worthy of a recommendation.

In a word: Highly Recommended

In this article we test GP(GPU) integrated graphics performance; please see our other articles on:

AMD Ryzen 2 Mobile 2500U Review & Benchmarks – Cache & Memory Performance

What is “Ryzen2” ZEN+ Mobile?

It is the long-awaited Ryzen2 APU mobile “Bristol Ridge” version of the desktop Ryzen 2 with integrated Vega graphics (the latest GPU architecture from AMD) for mobile devices. While on desktop we had the original Ryzen1/ThreadRipper – there was no (at least released) APU version or a mobile version – leaving only the much older designs that were never competitive against Intel’s ULV and H APUs.

After the very successful launch of the original “Ryzen1”, AMD has been hard at work optimising and improving the design in order to hit TDP (15-35W) range for mobile devices. It has also added the brand-new Vega graphics cores to the APU that have been incredibly performant in the desktop space. Note that mobile versions have a single CCX (compute unit) thus do not require operating system kernel patches for best thread scheduling/power optimisation.

Here’s what AMD says it has done for Ryzen2:

  • Process technology optimisations (12nm vs 14nm) – lower power but higher frequencies
  • Improvements for cache & memory speed & latencies (we shall test that ourselves!)
  • Multi-core optimised boost (aka Turbo) algorithm – XFR2 – higher speeds

Why review it now?

With Ryzen3 soon to be released later this year (2019) – with a corresponding Ryzen3 APU mobile – it is good to re-test the platform especially in light of the many BIOS/firmware updates, many video/GPU driver updates and not forgetting the many operating system (Windows) vulnerabilities (“Spectre”) mitigations that have greatly affected performance – sometimes for the good (firmware, drivers, optimisations) sometimes for the bad (mitigations).

In this article we test CPU Cache and Memory performance; please see our other articles on:

Hardware Specifications

We are comparing the top-of-the-range Ryzen2 (2700X, 2600) with previous generation (1700X) and competing architectures with a view to upgrading to a mid-range high performance design.

 

CPU Specifications AMD Ryzen2 2500U Bristol Ridge Intel i7 6500U (Skylake ULV) Intel i7 7500U (Kabylake ULV) Intel i5 8250U (Coffeelake ULV) Comments
L1D / L1I Caches 4x 32kB 8-way / 4x 64kB 4-way 2x 32kB 8-way / 2x 32kB 8-way 2x 32kB 8-way / 2x 32kB 8-way 4x 32kB 8-way / 4x 32kB 8-way Ryzen2 icache is 2x of Intel with matching dcache.
L2 Caches 4x 512kB 8-way 2x 256kB 16-way 2x 256kB 16-way 4x 256kB 16-way Ryzen2 L2 cache is 2x bigger than Intel and thus 4x larger than older SKL/KBL-U.
L3 Caches 4MB 16-way 4MB 16-way 4MB 16-way 6MB 16-way Here CFL-U brings 50% bigger L3 cache (6 vs 4MB) which may help some workloads.
TLB 4kB pages
64 full-way / 1536 8-way 64 8-way / 1536 6-way 64 8-way / 1536 6-way 64 8-way / 1536 6-way No TLB changes.
TLB 2MB pages
64 full-way / 1536 2-way 8 full-way  / 1536 6-way 8 full-way  / 1536 6-way 8 full-way  / 1536 6-way No TLB changes, same as 4kB pages.
Memory Controller Speed (MHz) 600 2600 (400-3100) 2700 (400-3500) 1600 (400-3400) Ryzen2’s memory controller runs at memory clock (MCLK) base rate thus depends on memory installed. Intel’s UNC (uncore) runs between min and max CPU clock thus perhaps faster.
Memory Speed (MHz) Max
1200-2400 (2667) 1033-1866 (2133) 1067-2133 (2400) 1200-2400 (2533) Ryzen2 now supports up to 2667MHz (officially) which should improve its performance quite a bit – unfortunately fast DDR4 is very expensive right now.
Memory Channels / Width
2 / 128-bit 2 / 128-bit 2 / 128-bit 2 / 128-bit All have 128-bit total channel width.
Memory Timing (clocks)
17-17-17-39 8-56-18-9 1T 14-17-17-40 10-57-16-11 2T 15-15-15-36 4-51-17-8 2T 19-19-19-43 5-63-21-9 2T Timings naturally depend on memory which for laptops is somewhat limited and quite expensive.
Memory Controller Firmware
2.1.0 3.6.0 3.6.4 Firmware is the same as on desktop devices.

Core Topology and Testing

As discussed in the previous articles (Ryzen1 and Ryzen2 reviews), cores on Ryzen are grouped in blocks (CCX or compute units) each with its own L3 cache – but connected via a 256-bit bus running at memory controller clock. However – unlike desktop/workstations – so far all Ryzen2 mobile designs have a single (1) CCX thus all the issues that “plagued” the desktop/workstation Ryzen designs do note apply here.

However, AMD could have released higher-core mobile designs to go against Intel’s H-line (beefed to 6-core / 12-threads with CFL-H) that would have likely required 2 CCX blocks. At this time (start 2019) considering that Ryzen3 (mobile) will launch soon that seems unlikely to happen…

Native Performance

We are testing native arithmetic, SIMD and cryptography performance using the highest performing instruction sets (AVX2, AVX, etc.). Ryzen2 mobile supports all modern instruction sets including AVX2, FMA3 and even more.

Results Interpretation: Higher rate values (GOPS, MB/s, etc.) mean better performance. Lower latencies (ns, ms, etc.) mean better performance.

Environment: Windows 10 x64, latest AMD and Intel drivers. 2MB “large pages” were enabled and in use. Turbo / Boost was enabled on all configurations.

Native Benchmarks AMD Ryzen2 2500U Bristol Ridge Intel i7 6500U (Skylake ULV) Intel i7 7500U (Kabylake ULV) Intel i5 8250U (Coffeelake ULV) Comments
CPU Multi-Core Benchmark Total Inter-Core Bandwidth – Best (GB/s) 18.65 [-21%] 16.81 18.93 23.65 Ryzen2 L1D is not as wide as Intel’s designs (512-bit) thus inter-core transfers in L1D are 20% slower.
CPU Multi-Core Benchmark Total Inter-Core Bandwidth – Worst (GB/s) 9.29 [=] 6.62 7.4 9.3 Using the unified L3 caches – both Ryzen2 and CFL-U manage the same bandwidths.
CPU Multi-Core Benchmark Inter-Unit Latency – Same Core (ns) 16 [-24%] 21 18 19 Within the same core (share L1D) Ryzen2 has lower latencies by 24% than all Intel CPUs.
CPU Multi-Core Benchmark Inter-Unit Latency – Same Compute Unit (ns) 46 [-23%] 61 54 56 Within the same compute unit (shareL3) Ryzen2 again yields 23% lower latencies.
CPU Multi-Core Benchmark Inter-Unit Latency – Different Compute Unit (ns) n/a n/a n/a n/a With a single CCX we have no latency issues.
While the L1D cache on Ryzen2 is not as wide as on Intel SKL/KBL/CFL-U to yield the same bandwidth (20% lower), both it and L3 manage lower latencies by a relatively large ~25%. With a single CCX design we have none of the issues seen on the desktop/workstation CPUs.
Aggregated L1D Bandwidth (GB/s) 267 [-67%] 315 302 628 Ryzen2’s L1D is just not wide enough – even 2-core SKL/KBL-U have more bandwidth and CFL-U has almost 3x more.
Aggregated L2 Bandwidth (GB/s) 225 [-29%] 119 148 318 The 2x larger L2 caches (512 vs 256kB) perform better but still CFL-U manages 30% more bandwidth.
Aggregated L3 Bandwidth (GB/s) 130 [-31%] 90 95 188 CFL-U not only has 50% bigger L3 (6 vs 4MB) but also somehow manages 30% more bandwidth too while SKL/KBL-U are left in the dust.
Aggregated Memory (GB/s) 24 [=]
21 21 24 With the same memory clock, Ryzen2 ties with CFL-U which means good bandwidth for the cores.
While we saw big improvements on Ryzen2 (desktop) for all caches L1D/L2/L3 – more work needs to be done: in particular the L1D caches are not wide enough compared to Intel’s CPUs – and even L2/L3 need to be wider. Most likely Ryzen3 with native wide 256-bit SIMD (unlike 128-bit as Ryzen1/2) will have twice as wide L1D/L2 that should be sufficient to match Intel.

The memory controller performs well matching CFL-U and is officially rated for higher DDR4 memory – though on laptops the choices are more limited and more expensive.

Data In-Page Random Latency (ns) 91.8 [4-13-32] [+2.75x] 34.6 [3-10-17] 27.6 [4-12-22] 24.5 As on desktop Ryzen1/2 in-page random latencies are large compared to the competition while L1D/L2 are OK but L3 also somewhat large.
Data Full Random Latency (ns) 117 [4-13-32] [-16%] 108 [3-10-27] 84.7 [4-12-33] 139 Out-of-page latencies are not much different which means Ryzen2 is a lot more competitive but still somewhat high.
Data Sequential Latency (ns) 4.1 [4-6-7] [-31%]
5.6 [3-10-11] 6.5 [4-12-13] 5.9 Ryzen’s prefetchers are working well with sequential access with lower latencies than Intel
Ryzen1/2 desktop issues were high memory latencies (in-page/full random) and nothing much changes here. “In-Page/Random pattern” (TLB hit) latencies are almost 3x higher – actually not much lower compared to “Full/Random pattern” (TBL miss) – which are comparable to Intel’s SKL/KBL/CFL. On the other hand “Sequential pattern” yields lower latencies (30% less) than Intel thus simple access patterns work better than complex/random access patterns.
Looking at the data access latencies’ graph for Ryzen2 mobile – we see the “in-page/random” following the “full/random” latencies all the way to 8MB block where they plateau; we would have expected them to plateau at a lower value. See the “code access latencies” graph below.
Code In-Page Random Latency (ns) 17.6 [5-9-25] [+14%] 13.3 [2-9-18] 14.9 [2-11-21] 15.5 Code latencies were not a problem on Ryzen1/2 and they are OK here, 14% higher.
Code Full Random Latency (ns) 108 [5-15-48] [+19%] 91.8 [2-10-38] 90.4 [2-11-45] 91 Out-of-page latency is also competitive and just 20% higher.
Code Sequential Latency (ns) 8.2 [5-13-20] [+37%] 5.9 [2-4-8] 7.8 [2-4-9] 6 Ryzen’s prefetchers are working well with sequential access pattern latency but not as fast as Intel.
Unlike data, code latencies (any pattern) are competitive with Intel though CFL-U does have lower latencies (between 15-20%) but in exchange you get a 2x bigger L1I (64 vs 32kB) which should help complex software.
This graph for code access latencies is what we expected to see for data: “in-page/random” latencies plateau much earlier than “full/random” thus “TLB hit” latencies being much lower than “TLB miss” latencies.
Memory Update Transactional (MTPS) 7.17 [-7%] 6.5 7.72 7.2 As none of Intel’s CPUs have HLE enabled Ryzen2 performs really well with just 7% less transactions/second.
Memory Update Record Only (MTPS) 5.66 [+5%] 4.66 5.25 5.4 With only record updates it manages to be 5% faster.

SiSoftware Official Ranker Scores

Final Thoughts / Conclusions

We saw good improvement on Ryzen2 (desktop/workstation) but still not enough to beat Intel and a lot more work is needed both on L1/L2 cache bandwidth/widening and memory latency (“in-page” aka “TBL hit” random access pattern) that cannot be improved with firmware/BIOS updates (AGESA firmware). Ryzen2 mobile does have the potential to use faster DDR4 memory (officially rated 2667MHz) thus could overtake Intel using faster memory – but laptop DDR4 SODIMM choice is limited.

Regardless of these differences – the CPU results we’ve seen are solid thus sufficient to recommend Ryzen2 mobile especially when at a much lower cost than competing designs. Even if you do choose Intel – you will be picking up a better design due to Ryzen2 mobile competition – just compare the SKL/KBL-U and CFL/WHL-U results.

We are looking forward to see what improvements Ryzen3 mobile brings to the mobile platform.

In a word: Recommended – with reservations

In this article we tested CPU Cache and Memory performance; please see our other articles on: