SiSoftware Sandra 20/21-R2x – Pushing AVX512

Note: We have updated R2 with R2x to fix an OpenCL GP-GPU bug discovered by Mark Poppin.

We are pleased to release R2x (version 31.25) update for Sandra 20/21 with the following changes:

Benchmark Support, Updates & Fixes

CPU Scientific Benchmark: greatly improved AVX512 GEMM (SGEMM, DGEMM) performance by using the larger 48kB L1D cache (on RKL and ICL-SP) and better prefetch. [note AVX2/FMA3 performance is also improved to some extent]
CPU Neural-Network Benchmark: reasonably improved AVX512 performance by using additional registers (on RKL and ICL-SP) and better prefetch. [note AVX2/FMA3 performance is also reasonably improved]
CPU Inter-Thread Efficiency Benchmark: reasonably improved AVX512 inter-thread transfers by using larger transfer blocks that take advantage of the larger 48kB L1D cache (on RKL and ICL-SP). [note AVX2/FMA3 performance is also improved to some extent]
Memory Bandwidth Benchmark: modestly improved AVX512 bandwidth transfers by using the larger 48kB L1D cache( on RKL and ICL-SP) and smarter prefetching. [note AVX2/FMA3 performance is also improved to some extent]

Benchmark	Old Sandra (RTM)	New Sandra (R2)	Old Sandra ( RTM)	New Sandra (R2)
Processor	RocketLake (RKL) i9-11900K		Skylake-X (SKL-X) i9-7900X
Memory Bandwidth Integer AVX512	41 GB/s	42 GB/s [+2%]	56.3 GB/s	56.4 GB/s [+1%]
Memory Bandwidth Float AVX512	40.3 GB/s	41.4 GB/s [+3%]	56.1 GB/s	56.5 GB/s [+1%]
Inter-Core Bandwidth AVX512	83.7 GB/s	84.5 GB/s [+1%]	83.3 GB/s	94.68 GB/s [+14%]
Inter-Core Latencies AVX512 (lower is better)	31ns (14.7, 32.2)	28.3ns (13.1, 29.4) [-9%]	77ns (15.7, 80.9)	53.6ns (14.3, 55.5) [-30%]
DGEMM Double AVX512	190 GFLOPS	284 GFLOPS [+49%]	196 GFLOPS	333 GFLOPS [+70%]
SGEMM Float AVX512	337 GFLOPS	586 GFLOPS [+74%]	599 GFLOPS	830 GFLOPS [+38%]

Please see our article Processor Multi-Core Efficiency: Inter-Thread Latency & Bandwidth Benchmarking for an in-depth look at the benchmark and an example of its results from a modern processor.

Hardware Support, Updates & Fixes

GP-GPU Processing OpenCL Benchmark: fixed kernel failures on most GP-GPUs.
All CPU Client Benchmarks: support Intel Pentium & Celeron (Atom “JasperLake”).
All CPU Server Benchmarks: preliminary support Intel Xeon (“IceLake-SP”).
All CPU Hybrid Benchmarks: optimised Intel Core/Atom Hybrid (“Lakefield”) performance.
All CPU Mobile Benchmarks: initial fixes for detection that fails to detect TDP/other power values on AMD Mobile Ryzen all series. Please contact us if you have such a processor and can help with troubleshooting.

Additional Support, Updates & Fixes

Branding update: all brand images have re-done in high resolution for modern displays.
SiSoftware Official Benchmark Ranker: graphics update with new logo and rosette for #1 (gold), #2 (silver), #3 (bronze) and lower ratings.
All websites: Cloudflare CDN (HTTPS/3) to reduce client load times and support times of high load that would overwhelm the Ranker servers and cause DDOS. Please contact us if you experience any issues.

And please, don’t forget small ISVs like ourselves in these very challenging times. Please buy a copy of Sandra if you find our software useful. Your custom means everything to us!

Reviews using Sandra 20/21:

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 Commercial (Pro/Biz/Eng/Ent)

Download Sandra Lite