Using GPU but it still use CPU

vision
1

i try to GPU on PyTorch after i formatted my computer (before i formatted and it worked)
then, i try this code

torch.cuda.is_available()

and output is

True

when i try to do like this

x = torch.rand(250, 250)
x = x.cuda()
print(x)
while 1:
    y = x*x

output

tensor([[0.7734, 0.6112, 0.9969,  ..., 0.3762, 0.1328, 0.0282],
        [0.7798, 0.9056, 0.4914,  ..., 0.0698, 0.9631, 0.7149],
        [0.1020, 0.7871, 0.2783,  ..., 0.6885, 0.8713, 0.1423],
        ...,
        [0.4408, 0.7382, 0.0786,  ..., 0.5183, 0.6337, 0.5479],
        [0.2822, 0.8666, 0.2791,  ..., 0.6299, 0.9231, 0.6004],
        [0.7444, 0.7096, 0.2371,  ..., 0.3882, 0.2971, 0.2674]],
       device='cuda:0')

you will see my tensor is cuda device. but when i see my GPU usage it’s not using.
it’s use CPU

Can you guys tell me why ?

ps. I have installed my VGA driver and CUDA 9.2 driver already
ps2. i used Jupyter Notebook(Anaconda) on WIndows
ps3. It worked before I formatted my computer

2

The workload might be a bit too small to see some utilization in nvidia-smi.
Change the size to torch.randn(2500, 2500) and you should see some GPU utilization.

3

i used Windows OS then i use task manager to check it. and when i try a big tensor it will notice “cuda memory error”.

4

Then it is way too big and you are overflowing the GPU memory. Try something in between

5

Correct me if I’m wrong, but 2500*2500 float32 values should take (2500*2500*4)/1024**2 = 23.84MB?
@Amornpat_Champa Is your GPU somehow already filled?

6

yep, you are right. It’s around 24MB

7

You’re right. but it’s still not working.

8

Do you get the error after pushing the tensor to the GPU?
Could you post the complete error message?

9

Is it possible that the GPU is not freeing memory fast enough when processing the while 1 loop?

10

i didn’t get any error message.

11

i don’t think so. because i use this code on another notebook and it’s work.

12

@adrianjav @ptrblck thank you both for your help.

13

How did you notice the “cuda memory error” without an error message?