You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Because the CUDA version with cpu as Context did already work so far, I decided to try that already. I set the number of threads to use to one (same problem appears when picking 2 threads, didn't test for other values).
uci
setoption name Context value cpu
setoption name Threads value 1
isready
go movetime 5000
As confirmed with both htop and the System Monitor, CrazyAra takes up 100% of all my cores when thinking of a move, but that really shouldn't happen if I only specify one thread.
The text was updated successfully, but these errors were encountered:
Yes, that phenomena probably needs further explanation. CrazyAra will only use one thread for the MCTS tree management and tree traversal if Threads is set to 1.
However, the MXNet library will require as many threads as there are available on the system.
You need to set MXNET_OMP_MAX_THREADS and maybe also MXNET_CPU_PRIORITY_NTHREADS and MXNET_CPU_NNPACK_NTHREADS to 1 in order to only use one thread for the neural network inference.
The neural network inference requires the majority of the computational effort, especially for deeper more complex models.
Because the CUDA version with
cpu
asContext
did already work so far, I decided to try that already. I set the number of threads to use to one (same problem appears when picking 2 threads, didn't test for other values).As confirmed with both htop and the System Monitor, CrazyAra takes up 100% of all my cores when thinking of a move, but that really shouldn't happen if I only specify one thread.
The text was updated successfully, but these errors were encountered: