本帖最后由 jerrytsao 于 2019-11-23 19:14 编辑

转总现在修身养性, 不方便上论坛, 帮贴

https://www.sisoftware.co.uk/2019/11/20/intel-core-i9-10980x-cascade-lake-review-benchmarks-cpu-18-core-36-thread-avx512/

10980XE 18C vs 3950X 16C vs 7900X 10C vs 9900K 8C

Native Dhrystone Integer (GIPS)        779 [+3%]        400        455        753        CSL-X is just 3% faster than Ryzen3.

Native Dhrystone Long (GIPS)        835 [+11%]        393        448        750        With a 64-bit integer workload – the gain is 11%.

Native FP32 (Float) Whetstone (GFLOPS)        459 [-1%]        236        262        464        With floating-point workload we have a tie.

Native FP64 (Double) Whetstone (GFLOPS)        379 [-4%]        196        223        393        With FP64 it is 4% slower than R3.

Native Integer (Int32) Multi-Media (Mpix/s)        2,341 [+25%]        985        1,430        1,873        In this vectorised integer test, AVX512 allows CSL-X a 25% win.

Native Long (Int64) Multi-Media (Mpix/s)        913 [+23%]        414        550        744        With a 64-bit AVX2 integer workload the gain is 23%.

Native Quad-Int (Int128) Multi-Media (Mpix/s)        12.92 [=]        6.75        9.58        12.98        This is a tough test using Long integers to emulate Int128 without SIMD it’s a tie.

Native Float/FP32 Multi-Media (Mpix/s)        2,676 [+36%]        914        1,740        1,970        In this floating-point vectorised test, CSL-X is 36% faster.

Native Double/FP64 Multi-Media (Mpix/s)        1,738 [+45%]        535        1,140        1,200        Switching to FP64 SIMD code, the gain is 45%.

Native Quad-Float/FP128 Multi-Media (Mpix/s)        56.4 [+21%]        23        38.7        46.5        In this heavy algorithm using FP64 to mantissa extend FP128 CSL-X is 21% faster.

Crypto AES-256 (GB/s)        33.9 [2.6x]        17.6        34        13        With AES/HWA support CSL-X wins due to 4-channels.

Crypto AES-128 (GB/s)        33.9 [2.6x]        17.6        34        13        No change with AES128.

Crypto SHA2-256 (GB/s)        33.5 [+17%]        12        26        28.6        Without SHA/HWA CSL-X still wins.

Crypto SHA1 (GB/s)                22.9        38                Less compute intensive SHA1 .

Crypto SHA2-512 (GB/s)                9        21                SHA2-512 is not accelerated by SHA/HWA.

Black-Scholes float/FP32 (MOPT/s)                276        344                With non vectorised workload.

Black-Scholes double/FP64 (MOPT/s)        497 [+61%]        238        277        308        Using FP64 CSL-X is 60% faster than Ryzen3.

Binomial float/FP32 (kOPT/s)                59.9        68.3                Binomial uses thread shared data thus stresses the cache & memory system.

Binomial double/FP64 (kOPT/s)        128 [+3%]        61.6        68        124        With FP64 code CSL-X is just 3% faster.

Monte-Carlo float/FP32 (kOPT/s)                56.5        257                Monte-Carlo also uses thread shared data but read-only thus reducing modify pressure on the caches.

Monte-Carlo double/FP64 (kOPT/s)        178 [-16%]        44.5        103        212        Switching to FP64 CSL-X is 16% slower.

SGEMM (GFLOPS) float/FP32                375        413                In this tough vectorised AVX2/FMA algorithm.

DGEMM (GFLOPS) double/FP64        240 [+45%]        209        212        165        With FP64 vectorised code, CSL-X is 45% faster.

SFFT (GFLOPS) float/FP32                22.3        28.6                FFT is also heavily vectorised but stresses the memory sub-system more.

DFFT (GFLOPS) double/FP64        22.07 [+2.6x]        11.21        14.6        8.56        With FP64 code, CSL-X is over 2x faster.

SNBODY (GFLOPS) float/FP32                557        638                N-Body simulation is vectorised but with more memory accesses.

DNBODY (GFLOPS) double/FP64        292 [-25%]        171        195        388        With FP64 code CSL-X is 25% slower.

Blur (3×3) Filter (MPix/s)        7,295 [+2.53x]        2,560        4,880        2,883        In this vectorised integer workload CSL-X is 2.5x faster.

Sharpen (5×5) Filter (MPix/s)        2,868 [+54%]        1,000        1,880        1,857        Same algorithm but more shared data still 54%.

Motion-Blur (7×7) Filter (MPix/s)        1,724 [+80%]        519        1,000        959        Same algorithm but even more data shared 80% faster.

Edge Detection (2*5×5) Sobel Filter (MPix/s)        2,285 [+44%]        827        1,500        1,589        Different algorithm but still vectorised workload 44% faster.

Noise Removal (5×5) Median Filter (MPix/s)        332 [+91%]        78        221        174        Still vectorised code again almost 2x faster.

Oil Painting Quantise Filter (MPix/s)        112 [+2.25x]        42.2        66.7        49.7        Even better improvement here of 2.25x

Diffusion Randomise (XorShift) Filter (MPix/s)        3,573 [2.37x]        4,000        3,084        1,505        With integer workload 2.5x faster.

Marbling Perlin Noise 2D Filter (MPix/s)        1,162 [+85%]        596        776        627        In this final test again with integer workload CSL-X is 85% faster.

结论

Thanks to AVX512 CSL-X manages to easily beat Ryzen3 in heavily vectorised algorithms (up to 50% faster) and also in memory-bandwidth heavy algorithms (due to its 4-channels memory sub-system). But, despite having 2 extra cores, on older AVX2/FMA we pretty much have a tie – not something we are used to see from Intel.

Also, the improvement over older SKL-X is exactly in-line with the increase in cores (18 vs 10 here) – thus there are no appreciable core improvements to boost performance. Without specific VNNI-accelerated algorithms, there is no point for SKL-X users to upgrade: you do get more cores for a lot less money but your hardware is also worth a lot less.

It shows how much Ryzen3 has improved (especially due to 256-bit width AVX2/FMA units) and ThreadRipper with 4-channels and even more cores (up to 32) and threads (up to 64) should nullify Intel's AVX512 benefit.