Pytorch for very old GPU

Prakhar Sharma

4/25/23, 3:40 AM

I've got Nvidia Quadro 2000 1GB from year 2010. From Nvidia's CUDA legacy GPUs list i can see it supports compute capability of 2.1.

I found a CUDA compatibility manual with a table of "CUDA Toolkit and Compatible Driver Versions" on page 4.

My driver version is 390.144 (from nvidia-smi), thus from the table CUDA 9.0 (9.0.76) should support the GPU.

Now I try to find an older version of Pytorch that supports CUDA 9.0. Fortunately, Pytorch 1.1.0 supports CUDA 9.0. Then I used this command to install this version of Pytorch conda install pytorch==1.1.0 torchvision==0.3.0 cudatoolkit=9.0 -c pytorch.

Unfortunately, torch.cuda.is_available() shows False. Am I doing something wrong? Btw this is the output of nvidia-smi command.

(base) hell@Dell-Precision-T1600:~$ nvidia-smi
Sat Dec 25 03:39:47 2021       
+---------------------------------------------+
| NVIDIA-SMI 390.144                Driver Version: 390.144                   |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  Quadro 2000         Off  | 00000000:01:00.0  On |                  N/A |
| 33%   61C    P0    N/A /  N/A |    383MiB /   963MiB |     24%      Default |
+-------------------------------+----------------------+----------------------+
                                                                               
+---------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |
|=============================================================================|
|    0       938      G   /usr/lib/xorg/Xorg                            93MiB |
|    0      1249      G   /usr/bin/gnome-shell                         159MiB |
|    0      3903      G   ...AAgAAAAAAAAACAAAAAAAAAA= --shared-files   125MiB |
|    0     17545      G   ...l/anaconda3/envs/pytorch_try/bin/python     1MiB |
+---------------------------------------------+

0 + 0

nvidia

cuda

gpu

ubfan1

4/25/23, 5:34 AM

Look at section 3.1 Hardware Support, table 2 of your cuda compatibility link. Your 2.1 compute capability (Fermi) doesn't even run driver 384. Try dropping back to CUDA 8 and the 367-375 driver.

cocomac

4/25/23, 5:59 AM

@ubfan1 I disagree. Plugging the OP's GPU into the [NVIDIA "Download Drivers" tool](https://www.nvidia.com/Download/index.aspx?lang=en-us) suggests version 390.147. That would put the OP at CUDA 9.0, as they are just short of the requirement for 9.1

Prakhar Sharma

4/25/23, 4:36 PM

@ubfan1 I tried downgrading and broke the display driver. Had black screen on boot. Then used chroot environment to uninstall nivida drivers and update grub.

ubfan1

4/25/23, 5:07 PM

I confirm my last driver used on my Quadro 1000 was the 390 driver, so that should work with you quadro 2000. Looks like Nvidia's older hardware web pages are not being maintained much, many dead links, like on our specific old Quadros.

Prakhar Sharma

4/25/23, 5:56 PM

Exactly, I don't know why CUDA 9.0 supported Pytorch is not recognizing my GPU.

Elon Musk

I sit in a Tesla and translated this thread with Ai:

EN: Pytorch for very old GPU

TH: Pytorch สำหรับ GPU เก่ามาก

RO: Pytorch pentru GPU foarte vechi

RU: Pytorch для очень старых GPU

VI: Pytorch cho GPU rất cũ

Post an answer

Most people don’t grasp that asking a lot of questions unlocks learning and improves interpersonal bonding. In Alison’s studies, for example, though people could accurately recall how many questions had been asked in their conversations, they didn’t intuit the link between questions and liking. Across four studies, in which participants were engaged in conversations themselves or read transcripts of others’ conversations, people tended not to realize that question asking would influence—or had influenced—the level of amity between the conversationalists.