Werner wrote: ↑Sun Sep 09, 2018 3:58 pm
I did not know, how strong these 6x64 net with LC0 is:
1 Lc0-1 9154 1CPU 3205 +436 111½1111111111101111 18.5/20
2 Scorpio 2.8 x64 1CPU 2769 -436 000½0000000000010000 1.5/20
perhaps I will make some more games with it...
sorry - wrong configuration with Scorpio here - I have to repeat this match.
Daniel Shawul wrote: ↑Sat Sep 08, 2018 8:30 am
I have made some significant work with GPU acceleration of NN evaluation in this version.
Multi-threaded batching as I have described in the programming section helps a lot and now scorpio
can be launched with upto 256 threads even on a 4-core machine and a single GPU and perform well.
I get 10kn/s on a 6x64 net on a P100 GPU.
Multi-threaded batching also helps tremendously on the CPU. On a 32-core machine, I am now getting 6.4kn/s
on a 6x64 net which is about 64% of eps that I got on a P100 GPU. I think batching helps in two ways -- a) minimizing
data transfer costs b/n CPU and GPU and b) minimizing the tensorflow call overhead. I think the latter is why the CPU
is benefiting from batching now while the GPU also benefits from (a) as well.
@Werner, the current scorpio-mcts+nn binary has issues with finishing mates (endgames). I haven't done any games before
I release. I have fixed that now along with fixing multi-core issues. I will upload this once I find workarounds for multi-GPU issue that
may take some time to fix ...
Quick question, Is there a way to specify which GPU to use. I have GPU0 and GPU1, I can tell it to use 1 or 2 GPU's but in the case of specifying which GPU to use I am unsure.
Thanks.
"Good decisions come from experience, and experience comes from bad decisions."
__________________________________________________________________
Ted Summers
I will add that feature in the next release but for now you can try to set the CUDA_VISIBLE_DEVICES environment variable for the scorpio process.
If you have 4 GPUs and you want to use /gpu:1 and /gpu:3 you just set CUDA_VISIBLE_DEVICES=1,3. Internally scorpio will still recognize these two devices as /gpu:0 and /gpu:1 because the numbering is local.
Hi 2.8.8 CPU version was working fine, downloaded latest egbbdll64 and epd and binaries for win 7 today . i start engine now type "go".
crashes with
Fault Module Name: StackHash_1dc2
Is it possible to compile a version for the 32 bit CPU armv7l Raspberry Pi? I managed to compile LC0 using openblas and can get up to 100 nps on 3 cores, but I’m having problems compiling your engine.
Daniel Shawul wrote: ↑Sun Sep 23, 2018 5:13 pm
If your CPU is older than 2011 (no AVX intructions) the new one won't work.
I will recompile and put up both of them.
Daniel Shawul wrote: ↑Sun Sep 23, 2018 5:13 pm
If your CPU is older than 2011 (no AVX intructions) the new one won't work.
I will recompile and put up both of them.
That will be great - thankyou!!
I have uploaded both now. The default supports sse2 intructions and is the one that should work for you.
Those with modern cpus can use the one in avx/ sub-folder.
Let me know if it works for you now.