在 Python 中使用 CUDA 进行 GPU 控制

Question

I want to use Tensorflow-gpu in Python. My PC has two GPUs. (GPU-A and GPU-B)

I want to specify the GPU to be used.

So when I run the program, I used the following command:

CUDA_VISIBLE_DEVICES = 1

The GPU to be used can be specified according to the value. Specifically, it was assigned as follows.

CUDA_VISIBLE_DEVICES = 0

I was able to use GPU-B.

On the other hand

CUDA_VISIBLE_DEVICES = 1

At that time, I was able to use GPU-A.

However, I cannot know in advance whether GPU-A or GPU-B corresponds to the value 0 or 1. I found a statement that "this index is assigned in descending order of computing power". However, GPU-A can calculate faster than GPU-B. I think that "indexes are assigned in order of the revision number (Compute Capability), not the calculation speed". In fact, GPU-B has higher compute capability than GPU-A.

If the hypothesis is correct, I have another question. If multiple GPUs with the same Compute Capability are used, how will they be assigned?

Answer 1

You can try running nvidia-smi in your terminal. The output you get shows you GPU ordering. For example:

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 410.104      Driver Version: 410.104      CUDA Version: 10.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce GTX 108...  Off  | 00000000:01:00.0 Off |                  N/A |
|  0%   52C    P8    21W / 250W |      2MiB / 11176MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+
|   1  GeForce GT 1030     Off  | 00000000:02:00.0  On |                  N/A |
| 41%   54C    P0    N/A /  30W |   1055MiB /  2001MiB |      6%      Default |
+-------------------------------+----------------------+----------------------+

Here, I have 2 GPUs:

1. GPU 0: GTX 1080Ti
2. GPU 1: GT 1030

You can then further use indexes in TensorFlow .