Not logged inRybka Chess Community Forum
Up Topic The Rybka Lounge / Computer Chess / RYZEN FRITZ CHESS BENCHMARKS ???
- - By ralunger (**) Date 2017-03-03 06:26
does anybody here have already ryzen cpus?
could you share your fritz benchmarks with SMT DISABLED? PLS??
- - By gsgs (***) Date 2017-03-06 14:54
Fritz Chessbench (MT, points)
22981 , Intel Core i7-6900K (8c/16t)
20495 , AMD Ryzen 7 1800X
20495 , AMD Ryzen 7 1800X SMT on
15484 , AMD Ryzen 7 1800X SMT off
vor 7 Stunden
Parent - By David Evans (****) Date 2017-03-07 06:19
The 6900k is double the price.

The 1800x even on early bios is doing very well indeed.
Parent - - By Lukas Cimiotti (Bronze) Date 2017-03-07 06:47
I made some little tests and posted them in a German forum (sorry, I'm too lazy to translate):
These results were made with only one run to depth 30/31 just to get an impression. The results make clear there is a problem with bmi2 compiles - they should be avoided. But overall chess performance isn't bad.
Parent - - By Geomusic (*****) Date 2017-03-16 20:58
You going for dual socket naples when they come out?
Parent - - By Lukas Cimiotti (Bronze) Date 2017-03-17 11:25
Probably yes, but depends on the costs.
Parent - By Carl Bicknell (*****) Date 2017-03-17 22:35
I'm starting to think Icelake might have some improvements in it if Intel feels threatened.
Parent - - By gsgs (***) Date 2017-03-17 12:52 Edited 2017-03-17 12:59
translation :

Ryzen Chess Performance - BMI2 Problem
For the comparison AMD Ryzen against Intel Skylake I have carried out the following tests:

Hardware Intel: Xeon E3-1245v5 on MSI C236A Workstation with 2x 8GB Samsung DDR4 ECC 2133
Hardware AMD: Ryzen 1800X on Asus Prime B350 plus with 2x 8 GB Crucial DDR4 2133
Both run on load on all logical cores at 3.7 GHz


Engine Intel AMD difference
Komodo 10.3 7,75 Mn/s 14,86 Mn/s 92%
asmFishW_2017_02_15_popcnt 10,60 Mn/s 19,21 Mn/s 81%
asmFishW_2017_02_15_bmi2 10,91 Mn/s 14,25 Mn/s 31%
Stockfish_8_x64_popcnt 8,36 Mn/s 16,51 Mn/s 97%
Stockfish_8_x64_bmi2 8,69 Mn/s 12,36 Mn/s 42%

For these results, an inaccuracy of a few percent is in each case, since chess engines running multi-threaded do not work deterministically.
There are different compiler options - the popcnt versions use the valuable instruction POPCNT from the instruction set SSE 4.2. With a command count how many farmers are on the field. The bmi2 versions also use the BMI2 instruction set. This should improve the performance actually - with Intel folds also, as one can see from the results. Ryzen seems to be a serious problem.

In the single-thread mode (since the Ryzen runs to 4.1 GHz), the following results are obtained:
Stockfish_8_x64_popcnt 1.591 Mn / s, Stockfish_8_x64_bmi2 1.156 Mn / s. These results are, by the way, extremely accurate

- in single-thread mode the results are deterministic. So 27.3% performance loss through the BMI2 option.

I myself wrote a small tool in C ++, which automatically sets affinities automatically optimally - regardless of the number of cores. I use this for years - and have it adapted yesterday, that it also SMT at AMD correctly named.
The chess engines are usually controlled by a GUI. A freeware GUI is e.g. Arena: Arena Chess GUI for Linux and Windows - Welcome to Arena
You can also start the engine by clicking on it and then enter text commands that correspond to the UCI protocol. A sequence of instructions, e.g. like this:

setoption name threads value 8
setoption name hash value 2048
go depth 30

Means: use 8 threads, 2048 MB hash, calculate the position (without input basic) up to a depth of 30 units. The result is then time **** (consumed milliseconds), nps (nodes per second - the required performance value), nodes (the total number of calculated nodes = positions) and score (evaluation of the position in cp = centipawns = 1 / 100 pawns

With my new Asrock AB350 Pro4 it is now possible to switch off SMT. With a 1700X there are the following values (everything asmFishW - the CPU runs under load to 3.5 GHz):

popcnt Version: 13,48 Mn/s
onhe SMT bmi2 Version: 9,41 Mn/s

so, it should be clear that it's no  SMT problem

AMD Ryzen 7 1800X, 16 Threads, all threads @3.7GHz, Asus Prime X370-Pro:
18.215.946nps, 8cp, 2.118.569.234n, g1f3 e7e6
18.307.576nps, 24cp, 1.232.887.103n, d2d4 e7e6
18.275.286nps, 6cp, 1.650.093.853n, d2d4 g8f6
18.279.639nps, 7cp, 2.532.077.382n, e2e4 e7e6
18.244.647nps, 14cp, 1.976.132.497n, d2d4 e7e6

Intel i7-7700K, 8 Threads, all threads @5.0GHz, Asus Prime Z270-A:
12.830.001nps, 5cp,, d2d4 g8f6
12.890.656nps, 17cp, 1.268.621.023n, d2d4 e7e6
12.839.685nps, 11cp, 1.334.903.602n, d2d4 e7e6
12.825.199nps, 10cp, 1.358.714.454n, d2d4 g8f6
12.834.243nps, 20cp, 1.790.710.651n, d2d4 g8f6

upgrade-bundle for old PCs :
C-Power-Bundle AMD Ryzen 7 1700. 16GB RAM DDR4 PC2400.
ASRock AB350 PRO4 AM4.
8x3,00GHz(Turbo 3,70Ghz) AM4.
EUR 659.52  including shipping
Parent - - By jpqy (**) Date 2017-03-17 14:52
@Lukas..have you tried the Base version..possible it will go even faster then popcnt version?!

Parent - - By Lukas Cimiotti (Bronze) Date 2017-03-17 19:45
The base version is slower.
Parent - By jpqy (**) Date 2017-03-17 20:21
Parent - By Geomusic (*****) Date 2017-06-18 02:09
Going for two AMD EPYC 7281s for $1200 looks like a nice price point for the 16c/32T cpu. In tandum $1200 bucks for 32C/64T! And I bet whatever board I put them in they will be at least as fast as a 2x e5-2697v4 and without the $2000/per cpu price tag LOL
Parent - - By cma6 (****) Date 2017-03-19 15:47
Hi gsgs:
  What does "SMT" mean?
Parent - By cma6 (****) Date 2017-03-20 02:13
Fritz Chessbench (MT, points)
22981 , Intel Core i7-6900K (8c/16t)
20495 , AMD Ryzen 7 1800X
20495 , AMD Ryzen 7 1800X SMT on
15484 , AMD Ryzen 7 1800X SMT off
vor 7 Stunden

Thanks for the SMP url.
It is unclear from gsgs' tests whether SMT helps chess performance since
20495 , AMD Ryzen 7 1800X
20495 , AMD Ryzen 7 1800X SMT on

How is SMT turned on: in the BIOS?
Up Topic The Rybka Lounge / Computer Chess / RYZEN FRITZ CHESS BENCHMARKS ???

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill