MyCaffe / NCCL

Windows version of NVIDIA's NCCL ('Nickel') for multi-GPU training - please use https://github.com/NVIDIA/nccl for changes.
Other
56 stars 10 forks source link

Is there any guide about installation? #1

Closed ziyigogogo closed 5 years ago

ziyigogogo commented 5 years ago

Hi there, Thank you so much for the repo. Since I met a nccl problem on my windows machine. But kinda confused by you repo how to use these things? Thanks again.

ZoroDerVonCodier commented 5 years ago

@ziyigogogo The installation is pretty simple - you just need to make sure the DLL is in the directory of the application that uses it. To see how we use the NCCL functions, please see https://github.com/MyCaffe/MyCaffe/blob/master/CudaDnnDLL/Cuda%20Files/nccl.cu

I hope this helps!

ZoroDerVonCodier commented 5 years ago

One other note - we have found that the NCCL works best when all GPU's used are ALL in TCC mode. For more about TCC mode, see https://docs.nvidia.com/gameworks/content/developertools/desktop/nsight/tesla_compute_cluster.htm

Not all GPU's support the TCC mode, but the Titan X, Titan Xp and Quadro RTX 8000 do. Running nvidia-smi (https://developer.nvidia.com/nvidia-system-management-interface) will tell you whether or not the GPU(s) that you have support TCC or not.

And for information on NCCL see https://devblogs.nvidia.com/fast-multi-gpu-collectives-nccl/

Hope this helps!

ziyigogogo commented 5 years ago

@ZoroDerVonCodier Thank you so for the links&notes. I will work on it.