Intel Iris Gen 12 (UHD 750) Rocketlake Review & Benchmarks – GPGPU Performance

What is “RocketLake”?

It is the desktop/workstation version of the true “next generation” Core (gen 10+) architecture  – finally replacing the ageing “Skylake (SKL)” arch and its many derivatives that are still with us (“CometLake (CML)”, etc.). It is a combination of the “IceLake (ICL)” CPU cores launched about a year go and the “TigerLake (TGL)” Gen12 XE graphics cores launched recently.

With the new core we get a plethora of new features – some previously only available on HEDT platform (AVX512 and its many friends), improved L1/L2 caches, improved memory controller and PCIe 4.0 buses. Sadly Intel had to back-port the older ICL (not TGL) cores to 14nm  – we shall have to wait for future (desktop) processors “AlderLake (ADL)” to see 10nm on the desktop…

  • 14nm+++ improved process (not 10nm)
  • Gen12 (Xe-LP) graphics (up to 96 EU in TGL graphics but here 32 EU only)
  • Transcode support for all major algorithms (e.g. HEVC/H.265 HDR/10-12bit, AV1 decode)
  • PCIe 4.0 (up to 32GB/s with x16 lanes) – 20 (16+4 or 8+8+4) lanes
  • Thunderbolt 3 (and thus USB 3.2 2×2 support @ 20Gbps) integrated

While ICL has already greatly upgraded the GP-GPU to gen 11 cores (and more than doubled to 64EU for G7), TGL upgrades them yet again to “XE”-LP gen 12 cores now all the way up to 96EUs. While again most features seem to be geared towards gaming and media (with new image processing and media encoders) – there should be a few new instructions for AI – hopefully provided by a OpenCL extension.

Again there is no FP64 support (!) while FP16 is naturally supported at 2x rate as before. BF16 should also be supported by a future driver. Int32, Int16 performance has reportedly doubled with Int8 now supported and DP4A accelerated.

We do hope to see more GPGPU-friendly features in upcoming versions now that Intel is taking graphics seriously. Perhaps with the forthcoming DG1 discrete graphics

GP-GPU (UHD 750, Xe-LP) Performance Benchmarking

In this article we test GP-GPU core performance; please see our other articles on:

Hardware Specifications

We are comparing the middle-range Intel integrated GP-GPUs with previous generation, as well as competing architectures with a view to upgrading to a brand-new, high performance, design.

Specifications Intel UHD 750 (32C, RKL RocketLake) Intel Iris XE ULV (96C, TGL TigerLake, i7 1165G7) Intel Iris Plus ULV (64C, ICL IceLake, i7 1065G7) Intel UHD 630 (24C, CFL-R CoffeeLake) Comments
Arch / Chipset EV12 / G1 EV12 / G7 EV11 / G7 EV9.5 / GT2 Gen 12 graphics – the latest.
Cores (CU) / Threads (SP) 256 / 32 [+33%] 768 / 96 64 / 512 24 / 192 33% more cores vs. CFL.
Speed (Min-Turbo)
1.3GHz [+8%]
1.2GHz 1.1GHz 1.2GHz Turbo speed has slightly increased.
Power (TDP) 125W [+25%] 28W 15W 95W TDP has increased 25% over CFL
ROP / TMU / 24 / 48 16 / 32 8 / 16 ROPs and TMUs likely increased.
Shared Memory
64kB
64kB 64kB 64kB Same shared memory.
Constant Memory
3.2GB 3.2GB 3.2GB 3.2GB No dedicated constant memory but large.
Global Memory 2x DDR4 3200Mt/s 128-bit
2x LP-DDR4X 4267Mt/s 128-bit
2x LP-DDR4X 3733Mt/s 128-bit 2x DDR4-3000Mt/s 128-bit Supports faster DDR4 memory.
Memory Bandwidth
42GB/s 42GB/s 58GB/s 42GB/s Highest (possible) bandwidth ever
L1 Caches 64kB 64kB 16kB 16kB L1 is much larger.
L3 Cache 3.8MB 3.8MB 3MB 512MB L3 has modestly increased.
Maximum Work-group Size
256×256 256×256 256×256 256×256 Same workgroup size
FP64/double ratio
No! No! No! Yes, 1/16x No FP64 support in current drivers!
FP16/half ratio
2x 2x 2x 2x Same 2x ratio
Price / RRP (USD)
$399 [-17%]
n/a n/a $479 Keen price, 17% lower!

Disclaimer

This is an independent article that has not been endorsed nor sponsored by any entity (e.g. Intel). All trademarks acknowledged and used for identification only under fair use. Errors and omissions excepted (E&OE).

The article contains only public information available elsewhere on the Internet and not provided under NDA or embargoed. At publication time, the products have not been directly tested by SiSoftware and thus the accuracy of the benchmark scores cannot be verified; however, they appear consistent and do not appear to be false/fake.

Native OpenCL Performance

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

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

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

Processing Benchmarks Intel UHD 750 (32C, RKL RocketLake) Intel Iris XE ULV (96C, TGL TigerLake, i7 1165G7) Intel Iris Plus ULV (64C, ICL IceLake, i7 1065G7) Intel UHD 630 (24C, CFL-R CoffeeLake) Comments
GPGPU Arithmetic Benchmark Mandel FP16/Half (Mpix/s) 1,570 [+43%] 4,680 2,300 1,100 Xe is 43% faster than EV9.5
GPGPU Arithmetic Benchmark Mandel FP32/Single (Mpix/s) 723 [+28%] 2,230 1,160 567 Standard FP32 is just 28% faster.
GPGPU Arithmetic Benchmark Mandel FP64/Double (Mpix/s) 35* [1/4x] 107* 61.6* 134 Without native FP64 support Xe is 1/4 the speed of EV9.5.
GPGPU Arithmetic Benchmark Mandel FP128/Quad (Mpix/s) 3.9* [1/2x] 11.9* 6.6* 7 Emulated FP128 is 1/2x speed.
Starting off, we see decent scaling with RKL’s XE32 28-43% faster than old EV9.5. Despite the higher TDP and still at 14nm it is still a decent improvement.

As with ICL and TGL, the lack of FP64 native support means any 64-bit floating point is 1/2-1/4x slower than even the ancient EV9.5 of CFL. For FP64 workloads – you’ll just have to use the CPU. What is strange is that Intel was the 1st to provide native 64-bit floating-point in consumer hardware and now it has removed it?

* Emulated FP64 through FP32.

GPGPU Crypto Benchmark Crypto AES-256 (GB/s) 2.9 [+2.9x] 8.86 2.27 1 Integer performance is 3x faster than EV9.5.
GPGPU Crypto Benchmark Crypto AES-128 (GB/s) 3.54 [+2.6x] 12.25 3 1.37 Nothing much changes when changing to 128bit.
GPGPU Crypto Benchmark Crypto SHA2-256 (GB/s) 7.38 [+2.4x] 22.39 6 3.1 Xe is over 2x faster than EV9.5.
GPGPU Crypto Benchmark Crypto SHA1 (GB/s) 13.34 [+70%] 34.57 12 7.85 With 128-bit Xe is 70% faster.
GPGPU Crypto Benchmark Crypto SHA2-512 (GB/s) 9 2.11 1.22 64-bit integer workload is also stellar.
With integer workloads, Xe32 is finally clearly faster than EV9.5 of CFL – as much as 3x and at least 70% in all tests. That is a pretty impressive performance, similar to what we saw in our TGL review – using Xe96. Naturally with just 32 EU it cannot reach the same performance but the result is impressive still.
GPGPU Finance Benchmark Black-Scholes float/FP16 (MOPT/s) 1,111 [-28%] 3,380 1,900 1,540 With FP16 see a regression.
GPGPU Finance Benchmark Black-Scholes float/FP32 (MOPT/s) 486 [1/2x] 1,730 1,260 987 With FP32 we see a bigger regression.
GPGPU Finance Benchmark Binomial half/FP16 (kOPT/s) 116 [+5%] 390 241 111 Binomial uses thread shared data thus stresses the memory system.
GPGPU Finance Benchmark Binomial float/FP32 (kOPT/s) 124 [+16%] 376 245 107 With FP32, Xe32 is just 16% faster.
GPGPU Finance Benchmark Monte-Carlo half/FP16 (kOPT/s) 470 [+57%] 1,430 616 300 Monte-Carlo also uses thread shared data but read-only.
GPGPU Finance Benchmark Monte-Carlo float/FP32 (kOPT/s) 496 [+64%] 1,420 646 303 With FP32 code Xe32 is 64% faster.
For financial FP32/FP16 workloads, Xe32 is sometimes a lot faster than old EV9.5 – but we also see some regressions – very likely due to drivers which will be fixed either by us or Intel. Elsewhere Xe32  is 5-64% faster than EV9.5 which is again pretty impressive – though naturally cannot match TGL’s 96EU GP-GPU.

No point to test or mention lack of native FP64 again – it’s not there. You are not going to be running 64-bit financial workloads on this GP-GPU.

GPGPU Science Benchmark HGEMM (GFLOPS) float/FP16 528 [+36%] 1,690 842 387 Xe32 starts well with 36% improvement.
GPGPU Science Benchmark SGEMM (GFLOPS) float/FP32 221 [+63%] 690 389 136 With FP32, Xe32 is 64% faster!
GPGPU Science Benchmark HFFT (GFLOPS) float/FP16 33.32 [-34%] 106 63 50.46 Again we see a regression here.
GPGPU Science Benchmark SFFT (GFLOPS) float/FP32 33.3 [+1%] 57.74 39.32 32.92 With FP32, Xe32 ties with EV9.5
GPGPU Science Benchmark HNBODY (GFLOPS) float/FP16 652 [+49%] 1,680 922 439 Xe32 is 50% faster here.
GPGPU Science Benchmark SNBODY (GFLOPS) float/FP32 338 [+30%] 970 522 260 With FP32, Xe32 is 30% faster.
On scientific algorithms (FP32 and FP16), Xe32 does much better and manages to be 30-63% faster than old EV9.5 of RKL. We do see one regression and one minor improvement but again that is an early driver issue.

Shall we mention lack of FP64 again? No we won’t.

GPGPU Image Processing Blur (3×3) Filter single/FP16 (MPix/s) 5,740 3,160 1,300
GPGPU Image Processing Blur (3×3) Filter single/FP32 (MPix/s) 1,348 [+2.35x] 5,310 1,370 574 In this 3×3 convolution algorithm, Xe32 is over 2x faster
GPGPU Image Processing Sharpen (5×5) Filter single/FP16 (MPix/s) 2,250 938 388
GPGPU Image Processing Sharpen (5×5) Filter single/FP32 (MPix/s) 479 [+3.74x] 1,470 281 128 Same algorithm but more shared data, Xe32 is 3.7x faster.
GPGPU Image Processing Motion Blur (7×7) Filter single/FP16 (MPix/s) 1,190 916 391
GPGPU Image Processing Motion Blur (7×7) Filter single/FP32 (MPix/s) 257 [+2x] 785 289 134 With even more data Xe32 is 2x faster.
GPGPU Image Processing Edge Detection (2*5×5) Sobel Filter single/FP16 (MPix/s) 1,760 866 369
GPGPU Image Processing Edge Detection (2*5×5) Sobel Filter single/FP32 (MPix/s) 481 [+3.79x] 1,310 317 127 Still convolution but with 2 filters – 3.8x faster.
GPGPU Image Processing Noise Removal (5×5) Median Filter single/FP16 (MPix/s) 18.14 24.38 16.45
GPGPU Image Processing Noise Removal (5×5) Median Filter single/FP32 (MPix/s) 18.24 [+64%] 36.26 24.13 11.1 Different algorithm Xe32 64% faster.
GPGPU Image Processing Oil Painting Quantise Filter single/FP16 (MPix/s) 31.43 21.42 15
GPGPU Image Processing Oil Painting Quantise Filter single/FP32 (MPix/s) 12.39 [=] 25.89 21.66 12.43 Without major processing, Xe32 just matches EV9.5.
GPGPU Image Processing Diffusion Randomise (XorShift) Filter single/FP16 (MPix/s) 2,420 1,460 1,270
GPGPU Image Processing Diffusion Randomise (XorShift) Filter single/FP32 (MPix/s) 1,304 [+6%] 2,590 1,540 1,230 This algorithm is 64-bit integer heavy and Xe32 is 6% faster.
GPGPU Image Processing Marbling Perlin Noise 2D Filter single/FP16 (MPix/s) 2,000 1,230 640
GPGPU Image Processing Marbling Perlin Noise 2D Filter single/FP32 (MPix/s) 1,109 [+45%] 1,330 820 764 One of the most complex and largest filters, Xe32 is 45% faster.
For image processing tasks, Xe32 seems to do best, with up to 4x better performance over old EV9.5; we saw similar improvement in TGL GP-GPU review. In any case for such tasks, upgrading to Xe32 will give you a huge boost. (fortunately no FP64 processing here)
GPGPU Memory Bandwidth Internal Memory Bandwidth (GB/s) 26 [-22%] 49 39 33.36 Xe32 really needs faster memory.
GPGPU Memory Bandwidth Upload Bandwidth (GB/s) 11.02 [-34%] 18.5 18.76 16.65 Uploads are slower.
GPGPU Memory Bandwidth Download Bandwidth (GB/s) 11.5 [-30%] 17.72 18.93 16.53 Download bandwidth is not much better.
It seems Xe32 really benefits from faster physical memory, with DDR4 unable to match high-speed LP-DDR4X despite its low power. With fast, e.g. 3200+ memory Xe32 is likely to do much better and we will see much better performance.

SiSoftware Official Ranker Scores

Final Thoughts / Conclusions

Summary: Recommended (~15% improvement over old EV9.5): 7/10

Once again Intel seems to be taking graphics seriously: for the 2nd time in a row we have a major graphics upgrade with Xe with big upgrades in EV cores (count), performance and bandwidth. It is lucky RKL has ended up with 14nm+++ Xe Gen 12 graphics cores and not Gen 11. As we saw in our TGL review, it can make a big difference.

But unlike top-end TGL APUs with 96 EUs, here we have just 32 EUs which despite much higher TDP (though at 14nm+++ not 10nm) cannot perform miracles, but ignoring a few preformance regressions it generally ends up much faster than old EV9.5 of CFL/CML – but all in all it ends up just 15% faster which is a pity.

However, this is still a core aimed at gamers and it does not provide much for GP-GPU; the improved integer performance is very much welcome – 3-times better (!) but few and specific instructions for AI only. Lack of FP64 makes it unsuitable for high-precision financial and scientific workloads; something that the old EV7-9 cores could do reasonably well (all things considered).

For integrated graphics, this is not a problem – not many people would expect integrated GPU core to run compute-heavy workloads; however, the lack of FP64 support is still jarring considering we’ve been used to having it in just about all other graphics architectures – including all the old Intel architectures.

It does seem that Xe32 and thus RKL like TGL before it really needs faster memory to perform much better and with improved drivers and faster memory we will see much better performance. These days Intel is releasing updated drivers regularly, fixing issues and adding features thus the future looks pretty bright.

Summary: Recommended 7/10

Please see our other articles on:

Disclaimer

This is an independent article that has not been endorsed nor sponsored by any entity (e.g. Intel). All trademarks acknowledged and used for identification only under fair use. Errors and omissions excepted (E&OE).

The article contains only public information available elsewhere on the Internet and not provided under NDA or embargoed. At publication time, the products have not been directly tested by SiSoftware and thus the accuracy of the benchmark scores cannot be verified; however, they appear consistent and do not appear to be false/fake.

Tagged , , , , , , , . Bookmark the permalink.

Comments are closed.