op12no2 wrote:
If I was to buy a video card just to train Leela, what would be recommend? I can get it through work; so may as well.
Hard to give specific recommendations, but at this point generally something NVIDIA GTX 1060, 1070, 1080 or 1080 Ti based, basically depending on how much you want to spend. Above the 1080 Ti and below the 1060 the price/performance curve drops off rapidly.
As for the different vendors out there, check some reviews for specific models you can get your hands on. Personally, I prefer a quiet card that's energy efficient over an overclocked one that runs 100Mhz faster, has a bunch of LEDs and is loud. (It adds up quickly with a bunch in my office!)
AMD cards will run fine if you have them lying around, and are well supported by Leela, but not necessarily by other deep learning based things. For that reason it's still hard to recommend them.
Gian-Carlo Pascutto wrote:
Hard to give specific recommendations, but at this point generally something NVIDIA GTX 1060, 1070, 1080 or 1080 Ti based, basically depending on how much you want to spend. Above the 1080 Ti and below the 1060 the price/performance curve drops off rapidly.
As for the different vendors out there, check some reviews for specific models you can get your hands on. Personally, I prefer a quiet card that's energy efficient over an overclocked one that runs 100Mhz faster, has a bunch of LEDs and is loud. (It adds up quickly with a bunch in my office!)
AMD cards will run fine if you have them lying around, and are well supported by Leela, but not necessarily by other deep learning based things. For that reason it's still hard to recommend them.
Dann Corbit wrote:
The last line of the report was this:
(1269/5310) KWG=32 KWI=2 MDIMA=8 MDIMC=8 MWG=64 NDIMB=8 NDIMC=8 NWG=16 SA=0 SB=0 STRM=1 STRN=1 VWM=8 VWN=2 0.7962 ms (10.5 GFLOPS)
How do I start up with those settings?
They're saved in the leelaz_opencl_tuning file and remembered.
All for naught, it is clear to me that my GPU is not being used.
The end-around I found on the net failed, and my back-off to an older driver is hung for an hour now.
I guess it is time to buy a nice Nvidia card.
Some numbers as references for the speed :
On my old GTX 750 Ti : 0.7 Kn/s (700 n/s)
On a GTX 1080 Ti : around 2.5 Kn/s
On the TCEC server (43 cores @2.8 GHz) : 5 Kn/s
2 live video to watch LC0 playing :
and :
Some more numbers on my CPUs i7@6*4GHz :
with 6 cores : 0.55 Kn/s
with 1 cores : 0.1 Kn/s
with 12 threads (with hyperthreading) : 0.7 Kn/s
Gian-Carlo Pascutto wrote:
I said "publicly available".
We have no way to test against Fineart. Of course, they are free to setup Leela Zero in a handicapped manner in a private match and claim superiority. This is an approach that has been proven successful in the past and is hence likely to be tried again.
I never run on CGOS. It's a bit like trying to measure your engine's improvement by playing in the ChessBase engine room, except that you get disconnected (and hence lose on time) more.
Vinvin wrote:
Some numbers as references for the speed :
On my old GTX 750 Ti : 0.7 Kn/s (700 n/s)
On a GTX 1080 Ti : around 2.5 Kn/s
On the TCEC server (43 cores @2.8 GHz) : 5 Kn/s
2 live video to watch LC0 playing :
and :
Some more numbers on my CPUs i7@6*4GHz :
with 6 cores : 0.55 Kn/s
with 1 cores : 0.1 Kn/s
with 12 threads (with hyperthreading) : 0.7 Kn/s
Clear view of my hardware on depth=26 (except for 1 thread, only d=25) :