Sandra Lite (Free/Eval) 2012.SP4a (18.47)
[more=Что нового..]SP4a for Sandra 2012 is ready and can be downloaded! It resolves a few issues and improves performance on AMD Bulldozer/future Piledriver CPUs:
- Multi-Media: enabled FMA4 Multi-Media code-path, thus improving Bulldozer performance by over 20% (versus AVX). FMA3 is also supported for Piledriver (and Haswell) which may be a bit faster still.
- Crypto: improved SHA256/SHA1 Bulldozer bandwidth by 50% by rolling back one SNB AVX optimisation. SNB/IVB scores are not significantly affected by this change. Similarly for AVX2, though nothing supports it yet.
- Memory/Cache Bandwidth, Memory/Cache Latency: enabled large-pages (2MB) on Bulldozer by reading 2MB/TLB correctly. Assuming you granted yourself "lock pages in memory", using 2MB pages improves both performance and reliability by minimising TLB misses when using large memory blocks.
- Memory Bandwidth, Cache Bandwidth: improved both FMA4 and FMA3 code-paths for better STREAM/Triad performance (versus AVX).
- Memory/Cache Latency: 32-bit/x86: Rolled back assembler > intrinsic change as the x86 compiler decided to generate non-optimal loop for the latency test. x64 compiler worked just fine.
- Fixes: Turbo multiplier detection on Bulldozer where P-States are not as expected (e.g. when overclocked manually or BIOS setting them up incorrectly).[/more]
http://files.almodi.org/sisoftware/san1847.exe
[more=Что нового..]SP4a for Sandra 2012 is ready and can be downloaded! It resolves a few issues and improves performance on AMD Bulldozer/future Piledriver CPUs:
- Multi-Media: enabled FMA4 Multi-Media code-path, thus improving Bulldozer performance by over 20% (versus AVX). FMA3 is also supported for Piledriver (and Haswell) which may be a bit faster still.
- Crypto: improved SHA256/SHA1 Bulldozer bandwidth by 50% by rolling back one SNB AVX optimisation. SNB/IVB scores are not significantly affected by this change. Similarly for AVX2, though nothing supports it yet.
- Memory/Cache Bandwidth, Memory/Cache Latency: enabled large-pages (2MB) on Bulldozer by reading 2MB/TLB correctly. Assuming you granted yourself "lock pages in memory", using 2MB pages improves both performance and reliability by minimising TLB misses when using large memory blocks.
- Memory Bandwidth, Cache Bandwidth: improved both FMA4 and FMA3 code-paths for better STREAM/Triad performance (versus AVX).
- Memory/Cache Latency: 32-bit/x86: Rolled back assembler > intrinsic change as the x86 compiler decided to generate non-optimal loop for the latency test. x64 compiler worked just fine.
- Fixes: Turbo multiplier detection on Bulldozer where P-States are not as expected (e.g. when overclocked manually or BIOS setting them up incorrectly).[/more]
http://files.almodi.org/sisoftware/san1847.exe
