TensorFlow GPU: is cudnn optional? Couldn't open CUDA library libcudnn.so
I installed the tensorflow-0.8.0 GPU version, tensorflow-0.8.0-cp27-none-linux_x86_64.whl. It says it requires CUDA toolkit 7.5 and CuDNN v4.
# Ubuntu/Linux 64-bit, GPU enabled. Requires CUDA toolkit 7.5 and CuDNN v4. For
# other versions, see "Install from sources" below.
However, I accidently forget to install CuDNN v4, but it works OK besides the error message, "Couldn't open CUDA library libcudnn.so". But it works and says, "Creating TensorFlow device (/gpu:0)".
msg without CuDNN
I tensorflow/stream_executor/dso_loader.cc:105] successfully opened CUDA library libcublas.so locally
I tensorflow/stream_executor/dso_loader.cc:99] Couldn't open CUDA library libcudnn.so. LD_LIBRARY_PATH: /usr/local/cuda/lib64:
I tensorflow/stream_executor/cuda/cuda_dnn.cc:1562] Unable to load cuDNN DSO
I tensorflow/stream_executor/dso_loader.cc:105] successfully opened CUDA library libcufft.so locally
I tensorflow/stream_executor/dso_loader.cc:105] successfully opened CUDA library libcuda.so.1 locally
I tensorflow/stream_executor/dso_loader.cc:105] successfully opened CUDA library libcurand.so locally
('Extracting', 'MNIST_data/train-images-idx3-ubyte.gz')
/usr/lib/python2.7/gzip.py:268: VisibleDeprecationWarning: converting an array with ndim > 0 to an index will result in an error in the future
chunk = self.extrabuf[offset: offset + size]
/home/ubuntu/TensorFlow-Tutorials/input_data.py:42: VisibleDeprecationWarning: converting an array with ndim > 0 to an index will result in an error in the future
data = data.reshape(num_images, rows, cols, 1)
('Extracting', 'MNIST_data/train-labels-idx1-ubyte.gz')
('Extracting', 'MNIST_data/t10k-images-idx3-ubyte.gz')
('Extracting', 'MNIST_data/t10k-labels-idx1-ubyte.gz')
I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:900] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
I tensorflow/core/common_runtime/gpu/gpu_init.cc:102] Found device 0 with properties:
name: GRID K520
major: 3 minor: 0 memoryClockRate (GHz) 0.797
pciBusID 0000:00:03.0
Total memory: 4.00GiB
Free memory: 3.95GiB
I tensorflow/core/common_runtime/gpu/gpu_init.cc:126] DMA: 0
I tensorflow/core/common_runtime/gpu/gpu_init.cc:136] 0: Y
I tensorflow/core/common_runtime/gpu/gpu_device.cc:755] Creating TensorFlow device (/gpu:0) -> (device: 0, name: GRID K520, pci bus id: 0000:00:03.0)
I tensorflow/core/common_runtime/gpu/pool_allocator.cc:244] PoolAllocator: After 1704 get requests, put_count=1321 evicted_count=1000 eviction_rate=0.757002 and unsatisfied allocation rate=0.870305
I tensorflow/core/common_runtime/gpu/pool_allocator.cc:256] Raising pool_size_limit_ from 100 to 110
I tensorflow/core/common_runtime/gpu/pool_allocator.cc:244] PoolAllocator: After 1704 get requests, put_count=1812 evicted_count=1000 eviction_rate=0.551876 and unsatisfied allocation rate=0.536972
I tensorflow/core/common_runtime/gpu/pool_allocator.cc:256] Raising pool_size_limit_ from 256 to 281
Later, I installed CuDNN, but I don't see the differences.
msg with CuDNN
I tensorflow/stream_executor/dso_loader.cc:105] successfully opened CUDA library libcublas.so locally
I tensorflow/stream_executor/dso_loader.cc:105] successfully opened CUDA library libcudnn.so locally
I tensorflow/stream_executor/dso_loader.cc:105] successfully opened CUDA library libcufft.so locally
I tensorflow/stream_executor/dso_loader.cc:105] successfully opened CUDA library libcuda.so.1 locally
I tensorflow/stream_executor/dso_loader.cc:105] successfully opened CUDA library libcurand.so locally
('Extracting', 'MNIST_data/train-images-idx3-ubyte.gz')
/usr/lib/python2.7/gzip.py:268: VisibleDeprecationWarning: converting an array with ndim > 0 to an index will result in an error in the future
chunk = self.extrabuf[offset: offset + size]
/home/ubuntu/TensorFlow-Tutorials/input_data.py:42: VisibleDeprecationWarning: converting an array with ndim > 0 to an index will result in an error in the future
data = data.reshape(num_images, rows, cols, 1)
('Extracting', 'MNIST_data/train-labels-idx1-ubyte.gz')
('Extracting', 'MNIST_data/t10k-images-idx3-ubyte.gz')
('Extracting', 'MNIST_data/t10k-labels-idx1-ubyte.gz')
I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:900] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
I tensorflow/core/common_runtime/gpu/gpu_init.cc:102] Found device 0 with properties:
name: GRID K520
major: 3 minor: 0 memoryClockRate (GHz) 0.797
pciBusID 0000:00:03.0
Total memory: 4.00GiB
Free memory: 3.95GiB
I tensorflow/core/common_runtime/gpu/gpu_init.cc:126] DMA: 0
I tensorflow/core/common_runtime/gpu/gpu_init.cc:136] 0: Y
I tensorflow/core/common_runtime/gpu/gpu_device.cc:755] Creating TensorFlow device (/gpu:0) -> (device: 0, name: GRID K520, pci bus id: 0000:00:03.0)
I tensorflow/core/common_runtime/gpu/pool_allocator.cc:244] PoolAllocator: After 1704 get requests, put_count=1321 evicted_count=1000 eviction_rate=0.757002 and unsatisfied allocation rate=0.870305
I tensorflow/core/common_runtime/gpu/pool_allocator.cc:256] Raising pool_size_limit_ from 100 to 110
I tensorflow/core/common_runtime/gpu/pool_allocator.cc:244] PoolAllocator: After 1704 get requests, put_count=1811 evicted_count=1000 eviction_rate=0.552181 and unsatisfied allocation rate=0.537559
I tensorflow/core/common_runtime/gpu/pool_allocator.cc:256] Raising pool_size_limit_ from 256 to 281
So what's differences with/without CuDNN?
Answer
cuDNN is used to speedup a few TensorFlow operations such as the convolution. I noticed in your log file that you're training on the MNIST dataset. The reference MNIST model provided with TensorFlow is built around 2 fully connected layers and a softmax. Therefore TensorFlow won't attempt to call cuDNN when training this model.
I'm not sure that TensorFlow will automatically fallback to a slower convolution algorithm when cuDNN isn't available. If it doesn't you can always disable the use of cuDNN by setting the TF_USE_CUDNN environment variable to 0 before running TensorFlow.