Tensorflowgpu version is not powerful enough

1. The most important factor may be the insufficient utilization of GPU. For example, my machine runs fashion MNIST, the utilization rate of GPU is only about 3%, and GPU is twice as slow as CPU.

2. certainly Only NVIDIA's graphics card supports CUDA and tensorflow.

3.

Check whether the local GPU supports the installation of tensorflow GPU
CUDA and cudnn are required to install the GPU version of tensorflow
therefore, it is necessary to check whether CUDA supports native GPU, and the computing power should be greater than 3.1
according to the official data provided by NVIDIA, gf920m has the computing power of 3.5, which just meets the requirements, and can be installed with GPU version
the specific parameters are shown in the figure below

4. Tensor flow uses a GPU by default (if any)
the specific call will be in the program
if you want to use multiple GPUs, modify the program accordingly
as for which part will be run on GPU, generally speaking, it is the part of model training, because this part of the program is calculation intensity
other parts can be run on GPU, but e to the long communication time between CPU and GPU, the program execution will be slower.

5. In this respect, there is no way to switch. We can only redo the system. Only in this way can we achieve a good effect. Relatively speaking, there is no problem.

6. Need to buy external graphics card adapter card! Can install external graphics card, any graphics card can be

7. How to run tensorflow GPU mode on PC
how to compile the GPU version tensorflow on MAC
basic use
use tensorflow, You have to understand tensorflow:
use graph to represent computing tasks.
execute graph in the context called session.
use tenor to represent data.
maintain state through variables.
use feed and feed to assign values to or obtain data from any arbitrary operation

8. Just install the GPU version of tensorflow. PIP install upgrade also needs to install the GPU acceleration package. Download the CUDA and cudnn installation packages. Note that the version numbers are CUDA 8.0 and cudnn 5.1 respectively. If all are installed successfully, GPU will be used automatically when running the program

9.

You need to see if your PC configuration is enough. The GPU mode of TF only supports N cards, and then the computing power is higher than 3.0. The following installation steps are as follows:

1. First, register your NVIDIA developer account and download CUDA and cudnn respectively

2. Confirm to prepare the GCC version and install the dependency library sudo apt get install freegl

3. Install CUDA

4. Decompress cudnn

5. Clone tensorflow source code, configure configuration

6. Compile and install

7. The last test