GPU - minimum requirements?

leonowski · September 15, 2020, 10:45pm

Hi,

I’m running the gpu-x3-beta container and I get 400 errors from the API when an image is posted. If I use the CPU version, it works fine and returns results as expected. But, the CPU version is slow for my hardware and I was hoping I could use the GPU.

Is there a minimum GPU or version of CUDA that is required for this? I’m running a Geforce 710 and I am able to run the nvidia cuda runtime through docker and get an output from nvidia-smi. But, I have to run version 10 and not version 11.

 docker run --runtime=nvidia --rm nvidia/cuda:10.1-runtime nvidia-smi
Tue Sep 15 22:37:11 2020
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 440.59       Driver Version: 440.59       CUDA Version: 10.2     |
|-------------------------------+----------------------+----------------------+
| 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 GT 710      Off  | 00000000:01:00.0 N/A |                  N/A |
| 40%   36C    P0    N/A /  N/A |      0MiB /  2001MiB |     N/A      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |
|=============================================================================|
|    0                    Not Supported                                       |
+-----------------------------------------------------------------------------+

So - is there a minimum GPU or CUDA version that is needed for this to work?

Zais · September 16, 2020, 8:35am

Hello, I tested the same config many times with a gt 710 with 2gb of ram, never worked.

leonowski · September 16, 2020, 8:51am

Thanks for chiming in and letting me know. Ok, I guess I’ll try another GPU. Cheers!

Zais · September 16, 2020, 8:59am

Welcome, but I am still looking for a solution, I am not an expert, I hope a solution /confirmation from deepstack team…, maybe it is possible to run it with gt710…

leonowski · September 16, 2020, 9:39am

Well, based on what you said, i swapped out the GT 710 with a spare GTX 970 that I had around. I put that in just now and it works well. So, it seems the minimum version GPU required is Maxwell. The GT 710 is the generation before it, Kepler.

My previous times were 3 seconds for object analysis. With GPU, it is now max about 300ms. WOW!

tinnov · October 12, 2020, 6:59am

Sun Oct 11 22:59:42 2020
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 450.66       Driver Version: 450.66       CUDA Version: 11.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  Quadro K2200        Off  | 00000000:00:07.0 Off |                  N/A |
| 42%   51C    P8     1W /  39W |    757MiB /  4043MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|    0   N/A  N/A      8125      C   python3                           754MiB |
+-----------------------------------------------------------------------------+

[GIN] 2020/10/12 - 05:31:37 | 200 |  250.343745ms |      172.17.0.1 | POST     /v1/vision/detection
[GIN] 2020/10/12 - 05:31:42 | 200 |  241.107745ms |      172.17.0.1 | POST     /v1/vision/detection
[GIN] 2020/10/12 - 05:32:43 | 200 |  254.230065ms |      172.17.0.1 | POST     /v1/vision/detection
[GIN] 2020/10/12 - 05:32:48 | 200 |  246.060867ms |      172.17.0.1 | POST     /v1/vision/detection
[GIN] 2020/10/12 - 05:32:53 | 200 |  240.459923ms |      172.17.0.1 | POST     /v1/vision/detection
[GIN] 2020/10/12 - 05:32:58 | 200 |  255.665306ms |      172.17.0.1 | POST     /v1/vision/detection
[GIN] 2020/10/12 - 05:33:03 | 200 |  242.005577ms |      172.17.0.1 | POST     /v1/vision/detection
[GIN] 2020/10/12 - 05:34:15 | 200 |   257.00856ms |      172.17.0.1 | POST     /v1/vision/detection
[GIN] 2020/10/12 - 05:34:20 | 200 |  240.914136ms |      172.17.0.1 | POST     /v1/vision/detection
[GIN] 2020/10/12 - 05:34:25 | 200 |  235.695141ms |      172.17.0.1 | POST     /v1/vision/detection
[GIN] 2020/10/12 - 05:34:30 | 200 |  242.854151ms |      172.17.0.1 | POST     /v1/vision/detection

I’m using a NVIDIA Quadro K2200 with deepquestai/deepstack:gpu-x3-beta. Works fine so far, but had stability problems with deepquestai/deepstack:gpu.

Got it off eBay for like $80 delivered. I picked this card specifically because it has a low power footprint and requires no extra power supply cables, plus being a Quadro there’s no hackery needed to pass it through to a VM.