Life In 19x19

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 ^^

lightvector wrote: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

How strong is KataGo with an AMD Threadripper 3990X?
~1x RTX 2060
~1x RTX 2070
~1x RTX 2080
~1x RTX 2080 Ti


I mean KataGo running only on this CPU (no GPUs).

From what I have read, for 2 GPUs, there are only slight improvements for more than 16 real CPU cores. Threadripper would only make sense for using 4+ GPUs (say, 3080 or 3090), in which case networking / storage throughput are more decisive and general acceleration might be far below linear (e.g., 3x for 4 GPUs or 5.5x for 8 GPUs). OC, some programs might have better parallelisation. Cooling is also a decisive factor: much space around each GPU is mandatory. A 2000W power supply might, or might not, be enough.

lightvector wrote about 2 GPUs and KataGo in https://www.lifein19x19.com/viewtopic.p ... 28#p259528:
"you just use both GPUs. [...] if you have multiple tasks instead of just one, you just give both GPUs different tasks in parallel and get a 2x throughput. In an MCTS search, you always have multiple tasks. Different GPUs can evaluate different nodes in the tree. [However, SLI is not useful.]"

goame wrote in https://www.lifein19x19.com/viewtopic.p ... 83#p258983:
"2x RTX 2080 Ti are ~2x faster than 1x RTX 2080 Ti."

Therefore, I consider building a PC with 1x or 2x Nvidia RTX 3080 (on a mainboard with two PCIe x16 slots being at least three slots apart) and a ca. 12 cores AMD cpu (I know, 16 cores would be better, but also a bit expensive).

Is KataGo the only program using multiple GPUs effectively or what other programs also do it well?

goame wrote: I mean KataGo running only on this CPU (no GPUs).

Seeing your edit, uhm, I have no idea...

Different question: for your 2x 2080 TI, which CPU do you use and to what percentage is it used under load?

goame wrote:

lightvector wrote: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

How strong is KataGo with an AMD Threadripper 3990X?
~1x RTX 2060
~1x RTX 2070
~1x RTX 2080
~1x RTX 2080 Ti


I mean KataGo running only on this CPU (no GPUs).

Ryzen Threadripper 3990X ~3.5 TFlops
GeForce GTX 1060 6GB ~4.4 TFlops
GeForce RTX 2060 ~5.2 TFlops

RobertJasiek wrote:

goame wrote: I mean KataGo running only on this CPU (no GPUs).

Seeing your edit, uhm, I have no idea...

Different question: for your 2x 2080 TI, which CPU do you use and to what percentage is it used under load?

6 core CPU and 33% under load.

And wrote:

goame wrote:

lightvector wrote: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

How strong is KataGo with an AMD Threadripper 3990X?
~1x RTX 2060
~1x RTX 2070
~1x RTX 2080
~1x RTX 2080 Ti


I mean KataGo running only on this CPU (no GPUs).

Ryzen Threadripper 3990X ~3.5 TFlops
GeForce GTX 1060 6GB ~4.4 TFlops
GeForce RTX 2060 ~5.2 TFlops

This is very interesting.
The 4990X which should be available in the next time, could be as strong as the RTX 2060.
5990X could be like one RTX 2080 Ti.
Maybe it starts to make sense to use GPUs and CPUs when running KataGo?

another source https://www.techpowerup.com/cpu-specs/r ... 990x.c2271
AMD Ryzen Threadripper 3990X - 13,209.6 GFLOPS

GeForce RTX 2080 Ti - 13,448 GFLOPS

3990x = 3 *2080Ti

costs

and besides that, it is not known what the difference will be in a particular application