40 games test : KG 1.5.0 v. KG 1.4.5

gpu : 1x1080 , network : g170e-b20c256x2-s5303129600-d1228401921 , 1s/move for both, same config file.

KG 1.5.0 wins 22-18 (55%) All games by resignation, no error, no duplicate game.

Probably not a very meaningful test, but ... KG goes on

Stats :

These games are probably on CUDA? it would be interesting to see also on OpenCL ("New in this release is support for FP16 tensor core GPUs in OpenCL, roughly doubling performance." https://github.com/lightvector/KataGo/releases/)

"Also as of this release, there is a pure-CPU implementation which can be compiled via -DUSE_BACKEND=EIGEN for cmake. There are no precompiled executables for it right now because the implementation is very basic..." https://github.com/lightvector/KataGo/releases/

who knows how, please do it!

Have you tried OpenCL on the iGPU? It should be way faster than Eigen

my AMD Athlon X4 950 and AMD Ryzen 5 processors do not have an integrated graphics card. the thing is, I want to use KataGo in the background while doing something else. but when OpenCL is working, something else is not comfortable to do, because slows down! when using cpu-only I can select for example only 2 cores. with OpenCL it doesn't work out for me

You should buy a strong RTX 3000 GPU.
This gives endless fun.

if there was extra money, I'd rather buy a T-50. and, no doubt, I would have received much more pleasure!

Then just run the processor as an OpenCL device, it will still be faster than the Eigen backend.

iopq is it really possible? how to do it?

I found one question on SO about this
https://askubuntu.com/questions/821341/ ... plications

if you're on windows I don't know

yes, windows. but AMD A8-5600K with integrated graphics

benchmark leela-zero-0.17-cpuonly-win64 elf_v2 AMD Athlon X4 950
--benchmark -v500 -t4 -w networkfile.gz
1m 11s
10.2 average depth, 19 max depth
396 non leaf nodes, 1.28 average children
506 visits, 176493 nodes, 505 playouts, 8 n/s

benchmark leela-zero-0.17-win64 elf_v2 GT 610
--benchmark -v500 -w networkfile.gz
1m 07s
8.1 average depth, 17 max depth
397 non leaf nodes, 1.29 average children
513 visits, 179993 nodes, 512 playouts, 10 n/s

low-power processor ~ low-power graphics card

"borrow_tensorflow" version: x1400 speed up from 1.5.0 (70% of libtensorflow backend). (https://github.com/lightvector/KataGo/i ... -670855071)

can someone compile? (windows)

i try install it i follow this video : https://www.youtube.com/watch?v=uYT-32lCK5k but the neural network i DL is a .zip not a .bin.gz do i need dezip it ? if yes they are some others files not just the .bin.gz, what i do ? (the inner of the .zip : https://zupimages.net/viewer.php?id=20/34/zboc.png)

And what i must do for run the benchmark it's not clear since he already do it before the video and just comment :/

Fine Mouche https://github.com/lightvector/KataGo#how-to-use

Fine Mouche, you've learned you need a .bin.gz file. And then you looked inside the zip and you found a .bin.gz file... so maybe just try that .bin.gz that you found?

Yeah, just unzip it. The other stuff inside the zip file is the same neural net in alternative formats which are useful to people for other purposes but you don't need to care about them.

For the benchmark, follow And's link. Or, if you consider yourself inexperienced enough about computers to be uncomfortable unzipping files, specifying file paths or other things that involve command lines and text, you can consider trying something like https://github.com/sanderland/katrain, which is written to be more graphical from the start and download and configure things for you.

v1.6.0 is out, which now does have a reasonably-performing CPU version, along with a major bugfix in OpenCL for users who has multiple NVIDIA GPUs where it would essentially not be able to use both GPUs in parallel. (@And)

https://github.com/lightvector/KataGo/releases

huge thanks !!! it did happen! on my computer there is a network 10 b, maxVisits = 500, the time per move is GT 610 (OpenCL) and AMD Athlon X4 950 is the same - 10 sec
I set numSearchThreads = 3 and it's great! with OpenCL the image slowed down, but now the beauty !!!

lightvector I have no words ... you are a great man!

in the CPU version the 40x384 network connected easily! (https://d3dndmfyhecmj0.cloudfront.net/g ... 892979.zip)
in OpenCL did not work at all. no way (GT 610). apparently this is due to the large amount of memory consumed. in OpenCL it is video memory

My message took 3-4 days to get approved so during this time i found the solution ^^

So Yes i just dezip and use the .bin.gz ^^

And for the benchmark the video autor answer me a todo list for what do in the cmd ^^