Running Leela and Fat Fritz on your notebook

[AdminX]

I have been think about doing this project.

https://en.chessbase.com/post/running-l ... r-notebook

Video:

[M ANSARI]

Yes I saw that and did some searching on Amazon for the Jetson Xavier ... apparently it is quite expensive at around $900. If you have a home PC that is already with a good GPU then probably remote engine makes more sense. I wonder though, what is the power of this GPU compared to say an RTX 2080 Ti?

[AdminX]

Yes I saw that and did some searching on Amazon for the Jetson Xavier ... apparently it is quite expensive at around $900. If you have a home PC that is already with a good GPU then probably remote engine makes more sense. I wonder though, what is the power of this GPU compared to say an RTX 2080 Ti?

This Amazon link has it at $400 USD:

It's the JETSON AGX which lists for $750 USD: https://www.nvidia.com/de-de/autonomous ... gx-xavier/

Could be useful if in an area with poor internet speeds.

[yurikvelo]

When it comes to running an engine for purposes such as opening preparation and game analysis, my laptop (with GTX 1650) tends to crash from overheat.

I have desktop version of mobile GPU GTX 1650 and it's pretty hot and noisy.
If laptop is equipped with GTX 1650 and have stability issues - cheap laptop cooling pad will probably help

[M ANSARI]

I think this is a nice idea but this just seems way under powered for Lc0. If you look at the specs of even the top end model it is at around 500 cuda cores and 64 tensor cores. That is about 8x less than you would get for an rtx 2080ti card and price is around $1000. An rtx 2070 super is half that price with maybe 6x the Cuda and Tensor cores. I guess if you have a really weak laptop with 1 or 2 cores then this could be better than SF on 2 cores, but it does seem that just buying a 2070 super card and having a remote connection might be a much better option. Of course if you have no internet connectivity (impossible to imagine today with almost everyone with a mobile with internet) then it might be a feasible option. The thing with Lc0 is that unless you have some good hardware to run it, it really can give some poor analysis in some positions.

[smatovic]

There are external GPU solutions via Thunderbolt out there, just google for it....

regarding performance:

http://talkchess.com/forum3/viewtopic.p ... 92#p824192

I was thinking of getting a 2 in 1 laptop with an external gpu (2080ti) connected to thunderbolt or a desktop with the 2080ti. Would there be a significant difference in performance for Lela?

I was thinking if I got the 2 in 1 then I wouldn’t need a desktop.

lc0 isn't much bottlenecked by PCIe bandwidth so I wouldn't expect significant different in performance. Maybe about 5%.

Anyway, the Jetson hack is cool.

--
Srdja

[AdminX]

I think this is a nice idea but this just seems way under powered for Lc0. If you look at the specs of even the top end model it is at around 500 cuda cores and 64 tensor cores. That is about 8x less than you would get for an rtx 2080ti card and price is around $1000. An rtx 2070 super is half that price with maybe 6x the Cuda and Tensor cores. I guess if you have a really weak laptop with 1 or 2 cores then this could be better than SF on 2 cores, but it does seem that just buying a 2070 super card and having a remote connection might be a much better option. Of course if you have no internet connectivity (impossible to imagine today with almost everyone with a mobile with internet) then it might be a feasible option. The thing with Lc0 is that unless you have some good hardware to run it, it really can give some poor analysis in some positions.

True, however my laptop comes with a weak MX150 GPU (384 Cuda Cores) and I don't think it plays that badly either. The NVIDIA Jetson Xavier NX has 384 Cuda Cores and 48 Tensor Cores, so I am sure it plays even better. My RTX 2070 Super has 2,560 CUDA cores and 320 tensor cores and requires even way more power but has a form factor that I don't find feasible as a mobile solution. Everyone has different needs of course, but for myself I don't care to be dependent on the cloud or the speed on an internet connection. Now I can justify buying the NX model but not it's big brother. Then again I love playing around with different computer technologies and SOC (System on Chip) projects, Plus currently I think Chessbase is the only one that gives you Engine Cloud abilities and I don't want to have to depend on their service in order to use a stronger GPU when on the road. Mind you I agree with every point you have made, I just feel differently about the internet dependency factor. As for Stockfish, well Stockfish is Stockfish and Leela is Leela. They think differently and find different moves in different positions, I see no reason not to have both or favor one over the other.

[M ANSARI]

Actually reading up more on the subject, it does seem it is possible to get a device that can run your GPU in an enclosure and have that connected to your laptop. That might be more cost effective as you can simply use an 2070 Super card and have 6x the performance for the same price bracket. You probably have to do some technical acrobatics to get it to work though ... mind you this also holds true for the other system as well.

[AdminX]

Part 2 was just published on Chessbase website.

https://en.chessbase.com/post/running-l ... notebook-2

"The new version continues to support lc0 running externally on a portable Xavier device. It also now supports all other UCI-compliant engines (Fat Fritz, Stockfish, Fritz, Komodo, Houdini, etc.). In v2, ChessBase can also remotely access any of the UCI-compliant engines that are launched from Windows or Linux servers through networks such as home WiFi or Internet."

[AdminX]

With the Windows /Linux Agents you no longer need to rely on the cloud, if you only wish to do test on your own private network now you can. Banksia provides this feature already, Chessbase should have implemented this also, but at least Chessbase users who want this can have it also via this Agent.





http://www.ezchess.org/jetson_v2/UserGuide.html

###############################################################################
#EngineName Port EngineExecutable EngineArguments
#
#EngineName: The engine folder name in C:\JetsonBackend\
# EngineName is a user-defined name associated with one engine.
# You can assign separate EngineNames for the same type of engine
# but with different configurations. For example,
# lc0-cuda points to lc0.exe from GitHub development,
# lc0-cuda-0.25.1 points to lc0.exe from stable build, and
# lc0-cuda-lite points to lc0.exe with lighter weights.
#
#Port: TCP listening port. Each EngineName must be assigned a different
# port. Please do not use 53350, which is reserved for management.
# Any number between 49152 and 65535 is valid.
#
#Executable: Actual executable file name for each EngineName
# It is possible for different EngineNames to have the same
# Executable.
#
#EngineArguments: Engine specific settings or options
#
#Note: EngineExecutable must not have spaces. For example, the original Fritz
# executable is "Fritz 17.exe”, so you have to change the file name by
# replacing the space with other characters like an underscore or a dash.
# Then write the changed name in this configuration file.
###############################################################################

###############################################################################
# Port number can be anything you like as long as each line has a unique
# number assigned to it. We start with 53352 because the 5, 3, 3, 5, 2 are
# the telephone digits corresponding to "leela".
#
# 53350: reserved for agent management, DO NOT use this one!
# 533xx: lc0 - leela chess zero
# 544xx: sf - stockfish
# 555xx: ff - fat fritz
# 566xx: fz - fritz
# 577xx: ko - komodo
# 588xx: ho - houdini
# 599xx, 600xx, 611xx, 622xx, ...... Attention: max TCP port number is 65535
###############################################################################

lc0-cuda 53352 lc0.exe --backend=cudnn-auto:--weights=256x20-t40-1541.pb.gz
lc0-cuda-lite 53353 lc0.exe --backend=cudnn-auto:--weights=591226.pb.gz
#lc0-cuda-gpu1 53354 lc0.exe --backend=cudnn-auto:--backend-opts=gpu=1:--weights=256x20-t40-1541.pb.gz

sf-bmi2 54452 stockfish_20011801_x64_bmi2
sf 54453 stockfish_20011801_x64

#ff-cuda 55552 lc0-fatfritz-cuda.exe --backend=cudnn-auto:--weights=FatFritz.weights
#ff-cuda-rtx 55553 lc0-fatfritz-cuda.exe --backend=cudnn-fp16:--weights=FatFritz.weights

#fz 56652 Fritz_17