AndrewGrant wrote: ↑Thu Jul 08, 2021 12:30 am
Wolfgang wrote: ↑Wed Jul 07, 2021 2:44 pm
CPU
- Intel i5-6600k @ 4,2 GHZ, 16GB RAM
- Intel i7-4770 @ 3,8GHZ, 16GB RAM
- Intel i5-4440 @ 3,3 GHZ, 16GB RAM
- AMD Ryzen 7-2700 @ 3,5GHZ, 16GB RAM
- AMD Ryzen 5-1600 @ 3,5GHZ, 8GB RAM
- AMD Ryzen 5-1500x @ 3,8GHZ, 8GB RAM
I might be wrong, but I beleive that even though the Ryzen zen=1 series (1500X/1600) have AVX2, they do not have 256-bit channels for the operations, and they are emulated with 2x128 bit lanes, basically making the AVX2 versions no better than their smaller vector SSE3/SSSE3/SSE4.X.
For example, AVX2 is not faster than AVX on my Ryzen 1950x, but is massively faster on my Ryzen 3700x. The same holds true for some Intel machines. So ... test everything with the 2700
Hi Andrew,
I am not in a position to discuss the technical details due to lacking knowledge...
But I tried all compiles of Ethereal 13.00 on all mentioned machines using "go nodes" command in console mode. I think it were 20 million nodes.
Here are the results/conclusions:
- fastest variant is the pext-avx2-compile on 6600k and 4770 intel machines. 6600k is a bit faster due to to higher clock speed (4770 cannot be overclocked that much, its without "k")
- I disagree that the 2700 (without "x"!) is faster/fits better for NN-engines than the Ryzen 1xxx series. NPS are fairly equal.
- Maybe a Ryzen 2700 (or higher) with "x" will be considerably faster, because it can be better overclocked. Maybe one can overclock my Ryzen 2700 for more than 10%, but I look at stability first, not at ultimate performance.
So ... test everything with the 2700
No, I'll take 6600k and 4770 for Ethereal and other TOP engines...