DIGITS: Check failed: status == CUDNN_STATUS_SUCCESS (3 vs. 0) CUDNN_STATUS_BAD_PARAM

I have installed DIGITS 2.0RC using the webinstaller. I have installed CUDA 7.0.28. I am following the instructions to run MNIST dataset here: https://github.com/NVIDIA/DIGITS/blob/master/docs/GettingStarted.md#creating-a-dataset . When running LeNet network I get the following error:

Setting up mnist
Cannot create cuDNN handle. cuDNN won't be available.
Opened lmdb /home/alex/digits/digits/digits/jobs/20150901-134106-de89/train_db
Decoding Datum
output data size: 64,1,28,28
Top shape: 64 1 28 28 (50176)
Top shape: 64 (64)
Creating layer scale
Creating Layer scale
scale <- data
scale -> scale
Setting up scale
Cannot create cuDNN handle. cuDNN won't be available.
Top shape: 64 1 28 28 (50176)
Creating layer conv1
Creating Layer conv1
conv1 <- scale
conv1 -> conv1
Setting up conv1
Check failed: status == CUDNN_STATUS_SUCCESS (3 vs. 0)  CUDNN_STATUS_BAD_PARAM

I have seen others reporting about this issue but I have not seen an answer yet. Is it possible that CUDA has not been installed properly? How can verify that CUDNN is installed? Or is it perhaps some configuration I have to perform to the network before running? Thanks.

About this issue

Original URL
State: closed
Created 9 years ago
Comments: 21 (7 by maintainers)

Most upvoted comments

I also got this problem and it is just because my GPU is out of use, close the apps that are using GPU and make some memory for this, it can run.

CCChenfei on Mar 20, 2017

I’ll close this issue now as we have discovered the source of your problem. You have three options going forward:

Run digits-devserver with the --config flag and select “N” for 'None` in the GPUs section - that should let you run in CPU-only mode.
Build Caffe from source with CUDA support (but without cuDNN) - that might work, no promises (see here).
Upgrade your graphics card.

lukeyeager on Sep 2, 2015

Ah, there’s your problem. The GeForce GT 630 is a Fermi card, with compute capability 2.1. https://en.wikipedia.org/wiki/CUDA#Supported_GPUs

Section 2.5 of the cuDNN v3 User Guide says:

2.5. Requirements cuDNN v3.0 supports NVIDIA GPUs of compute capability 3.0 and higher and requires an NVIDIA Driver compatible with CUDA Toolkit 7.0.

Not that I would expect you to have read that. I’ll see if I can fix Caffe’s error reporting or at least improve their documentation.

lukeyeager on Sep 2, 2015