TalkChess.com

kasinp wrote: I should rephrase: the performance will be close, but the GPU will achieve it without breaking sweat (it will never exceed 80% utilization), in my case it doesn't even go above silent mode. OTOH, the i9 will behave like a decent space heater to achieve the similar level of play. In that sense I consider the GPU to be a hands-down better option.
PK

While running in GPU does LCZero use the GPU memory too or just the system RAM?

kasinp wrote:

kasinp wrote:

Milton wrote:I am getting a new computer around May 3 (a Dell Alien Area 51) and I would like to participate in this project.
Based on this configuration, would it be best for me to download the GPU file or the CPU file?

Intel® Core™ i9 7980XE
NVIDIA® GeForce® GTX 1080 with 8GB GDDR5X
32GB Dual Channel HyperX™ DDR4 XMP at 2933MHz

I would say GPU hands down. In my case a 1080 ti outperforms the CPU version running on a 14-core Xeon at 3.00GHz. Seems it really prefers the GPU architecture.
PK

I should rephrase: the performance will be close, but the GPU will achieve it without breaking sweat (it will never exceed 80% utilization), in my case it doesn't even go above silent mode. OTOH, the i9 will behave like a decent space heater to achieve the similar level of play. In that sense I consider the GPU to be a hands-down better option.

PK

OK. So I will download both files, but I will only run the GPU file unless we get cold weather here.

Jhoravi wrote:

kasinp wrote: I should rephrase: the performance will be close, but the GPU will achieve it without breaking sweat (it will never exceed 80% utilization), in my case it doesn't even go above silent mode. OTOH, the i9 will behave like a decent space heater to achieve the similar level of play. In that sense I consider the GPU to be a hands-down better option.
PK

While running in GPU does LCZero use the GPU memory too or just the system RAM?

I cannot say whether it uses the GPU memory or not, though presume it does, but it definitely uses the system RAM as well, even when running on the GPU.

Is there a command line flag or something to tell it to use the GPU?

You can run lc0 with --tune-only,
that will output a list of your OpenCL devices with OpenCL IDs,
then you can select the device with --gpu arg and run --tune-only and --full-tuner,
that will try a thousands of configuration options and create a config file for your device.

e.g. --
Srdja

It does not see my GPU at all.

I have no file weights.txt so I copied network.txt to weights.txt

G:\chess\LCZero>lczero --tune-only
Using 2 thread(s).
Detecting residual layers...v1...64 channels...6 blocks.
Initializing OpenCL.
Detected 1 OpenCL platforms.
Platform version: OpenCL 2.0 AMD-APP (1912.5)
Platform profile: FULL_PROFILE
Platform name: AMD Accelerated Parallel Processing
Platform vendor: Advanced Micro Devices, Inc.
Device ID: 0
Device name: Intel(R) Core(TM) i7-2600 CPU @ 3.40GHz
Device type: CPU
Device vendor: GenuineIntel
Device driver: 1912.5 (sse2,avx)
Device speed: 3392 MHz
Device cores: 8 CU
Device score: 520
Selected platform: AMD Accelerated Parallel Processing
Selected device: Intel(R) Core(TM) i7-2600 CPU @ 3.40GHz
with OpenCL 2.0 capability.
Loaded existing SGEMM tuning.

G:\chess\LCZero>lczero --gpu 0 --tune-only --full-tuner
Using 2 thread(s).
Detecting residual layers...v1...64 channels...6 blocks.
Initializing OpenCL.
Detected 1 OpenCL platforms.
Platform version: OpenCL 2.0 AMD-APP (1912.5)
Platform profile: FULL_PROFILE
Platform name: AMD Accelerated Parallel Processing
Platform vendor: Advanced Micro Devices, Inc.
Device ID: 0
Device name: Intel(R) Core(TM) i7-2600 CPU @ 3.40GHz
Device type: CPU
Device vendor: GenuineIntel
Device driver: 1912.5 (sse2,avx)
Device speed: 3392 MHz
Device cores: 8 CU
Device score: 520
Selected platform: AMD Accelerated Parallel Processing
Selected device: Intel(R) Core(TM) i7-2600 CPU @ 3.40GHz
with OpenCL 2.0 capability.

Started OpenCL SGEMM tuner.
RNG seed: 0xb1e44da6 (thread: 3568476555)
Will try 5238 valid configurations.
(1/5238) KWG=16 KWI=2 MDIMA=8 MDIMC=16 MWG=16 NDIMB=8 NDIMC=8 NWG=16 SA=0 SB=0 STRM=0 STRN=0 VWM=1 VWN=1 0.7277 ms (2.9 GFLOPS)
(4/5238) KWG=32 KWI=2 MDIMA=8 MDIMC=16 MWG=64 NDIMB=8 NDIMC=16 NWG=16 SA=0 SB=0 STRM=0 STRN=0 VWM=1 VWN=1 0.6078 ms (3.5 GFLOPS)
(5/5238) KWG=16 KWI=2 MDIMA=16 MDIMC=8 MWG=32 NDIMB=8 NDIMC=8 NWG=16 SA=0 SB=0 STRM=0 STRN=0 VWM=1 VWN=1 0.4344 ms (4.8 GFLOPS)
(78/5238) KWG=32 KWI=2 MDIMA=8 MDIMC=16 MWG=64 NDIMB=8 NDIMC=8 NWG=16 SA=0 SB=0 STRM=0 STRN=0 VWM=2 VWN=1 0.3680 ms (5.7 GFLOPS)
(107/5238) KWG=32 KWI=8 MDIMA=16 MDIMC=16 MWG=64 NDIMB=8 NDIMC=8 NWG=16 SA=0 SB=0 STRM=0 STRN=0 VWM=2 VWN=1 0.3461 ms (6.1 GFLOPS)
(127/5238) KWG=32 KWI=2 MDIMA=8 MDIMC=8 MWG=32 NDIMB=8 NDIMC=16 NWG=16 SA=0 SB=0 STRM=0 STRN=0 VWM=4 VWN=1 0.3450 ms (6.1 GFLOPS)
(129/5238) KWG=32 KWI=2 MDIMA=16 MDIMC=8 MWG=64 NDIMB=8 NDIMC=8 NWG=16 SA=0 SB=0 STRM=0 STRN=0 VWM=4 VWN=1 0.3315 ms (6.3 GFLOPS)
(133/5238) KWG=32 KWI=2 MDIMA=8 MDIMC=8 MWG=64 NDIMB=16 NDIMC=8 NWG=16 SA=0 SB=0 STRM=0 STRN=0 VWM=4 VWN=1 0.2737 ms (7.7 GFLOPS)
(519/5238) KWG=16 KWI=2 MDIMA=8 MDIMC=8 MWG=64 NDIMB=8 NDIMC=8 NWG=16 SA=0 SB=0 STRM=1 STRN=0 VWM=8 VWN=1 0.2609 ms (8.0 GFLOPS)

It does not see my GPU at all.

Yes, it lists only the CPU.

On Linux, the command "lspci | grep VGA" should list your GPU as PCI device.

Install and/or run the command "clinfo",
that should list all of your installed OpenCL devices.

Linux distros use usually the open source driver "Nouveau" for Nvidia devices,
so you have to install the proprietary Nvidia driver packages from your repository
(driver, libopencl, opencl-icd) to get OpenCL support.

I have no file weights.txt so I copied network.txt to weights.txt

G:\chess\LCZero>lczero --tune-only
Using 2 thread(s).
Detecting residual layers...v1...64 channels...6 blocks.

64 channels and 6 blocks, looks like you are running still version 0.5. or 0.4.

--
Srdja

G:\chess\LCZero>lczero --tune-only

Ah, just see, you are running windows,
so maybe a gpu driver update may help?

Afaik LC0 needs OpenCl >= 1.2, not sure if lower versions will work.

--
Srdja

Albert Silver wrote:

Jhoravi wrote:

kasinp wrote: I should rephrase: the performance will be close, but the GPU will achieve it without breaking sweat (it will never exceed 80% utilization), in my case it doesn't even go above silent mode. OTOH, the i9 will behave like a decent space heater to achieve the similar level of play. In that sense I consider the GPU to be a hands-down better option.
PK

While running in GPU does LCZero use the GPU memory too or just the system RAM?

I cannot say whether it uses the GPU memory or not, though presume it does, but it definitely uses the system RAM as well, even when running on the GPU.

GPU version 0.7 using 1080 ti also keeps 2-3 threads on my 14-core Xeon busy. During 10-15 min. analysis these typically allocate 3-4GB or RAM.

PK

I downloaded the GPU version, so I would assume it would use my GPU.
But right now I see LCZero using between 13 and 90% of my CPU.

It was not like that when I first started it. My CPU was quiet for some time.
I wonder if it downloads the CPU version by itself and then runs that.
What I see does not make sense to me.

Even the gpu version used the cpu for MCTS playout. (Default is one thread I think).
And as you have seen there are separate executables for gpu and cpu. GPU used opencl and cpu openblas IIRC