This is essentialy the config parts I needed to add to
/etc/nixos/configuration.nix in order to have CUDA. Particularly note the
SystemD unit, this creates the /dev/nvidia0, /dev/nvidia1,
/dev/nvidiactl files.
This directory contains all the example CUDA code from NVIDIA's CUDA Toolkit,
and a nix expression. Once your system is working (try testing with
nvidia-smi,) go into that directory, run:
nix-build default.nix -A examplecuda
It will tell you in there where it put it, at which point you can run those programs:
[grahamc@sc:/nix/store/a0v4qnwyr9ignk1mygncrk12cqarc58a-example-cuda/bin]$ ls
alignedTypes MonteCarloMultiGPU
asyncAPI newdelete
bandwidthTest p2pBandwidthLatencyTest
batchCUBLAS ptxjit
binomialOptions quasirandomGenerator
[grahamc@sc:~/nixos-cuda-k40]$ /nix/store/a0v4qnwyr9ignk1mygncrk12cqarc58a-example-cuda/bin/MonteCarloMultiGPU
/nix/store/a0v4qnwyr9ignk1mygncrk12cqarc58a-example-cuda/bin/MonteCarloMultiGPU Starting...
Using single CPU thread for multiple GPUs
MonteCarloMultiGPU
==================
Parallelization method = streamed
Problem scaling = weak
Number of GPUs = 2
Total number of options = 16384
Number of paths = 262144
main(): generating input data...
main(): starting 2 host threads...
main(): GPU statistics, streamed
GPU Device #0: Tesla K40c
Options : 8192
Simulation paths: 262144
GPU Device #1: GeForce GT 740
Options : 8192
Simulation paths: 262144
Total time (ms.): 344.402008
Note: This is elapsed time for all to compute.
Options per sec.: 47572.312637
main(): comparing Monte Carlo and Black-Scholes results...
Shutting down...
Test Summary...
L1 norm : 4.836368E-04
Average reserve: 11.301252
NOTE: The CUDA Samples are not meant for performance measurements. Results may vary when GPU Boost is enabled.
Test passed