CUDNN 6.1 is not available for download on https://developer.nvidia.com/rdp/cudnn-archive

[Kew]

CUDNN 6.1 (version listed in README.md) is not available for download on https://developer.nvidia.com/rdp/cudnn-archive

Is there another version I can use that would work?

[cbovar]

I have just tried with cuDNN v6.0 (April 27, 2017), for CUDA 8.0 and the unit tests pass. I will update REAME.md because 6.1 doens't seem to appear on nvidia website.

I need to update to Managed Cuda 9.1 / cuDNN 7.1

[Kew]

Thank you. I will try cuDNN v6.0 (April 27, 2017), for CUDA 8.0.

[Kew]

I think it is working... but not sure if I am comparing things right.

MnistDemo vs MnistDemo.GPU - GPU is getting through more examples but is converging slower than CPU version.

GPU
Epoch #7
Loss: 2.136335 Train accuracy: 28% Test accuracy: 38%
Example seen: 70656 Fwd: 0.27ms Bckw: 0.01ms Updt: 0ms

CPU
Epoch #1
Loss: 1.70971740924065 Train accuracy: 51% Test accuracy: 53%
Example seen: 680 Fwd: 16.44ms Bckw: 63.97ms

[cbovar]

There are some small differences between MnistDemo CPU and GPU.

• float vs double
• different batch size (20 vs 1024)
• GPU version doesn't use L2 regularization and Momentum (SgdTrainer)

You should get similar results If you set all those parameters to the same values. GPU should be faster if batchsize is big enough.