OpenNMT

Suggestion for Graphic cards?

Hello,

I’m looking in buying some serious graphic card(s). As I have became a serious user of Google Colab… I need something a little bit faster.

I was looking into the Nvidia Titan RTX, but honestly, I’m not too sure what parameters of the cards I should focus on to train NLP?

I don’t know if anyone would have any suggestion. I would say that I don’t have a clear budget but something around 2k to 3k (US) would be reasonable.

Best regards,
Samuel Lacombe

1 Like

Bigger VRAM means you can train bigger models (or models with longer sentence lengths). A Transformer-base model just needs 8GB, A Transformer-big needs ~16GB.

Obviously 1-2 3090 is currently the best, but they are overpriced; you could have bought 2 at MSRP.

4 Likes

I agree with James, 3090 is an absolute beast, but a few notes:

If you plan to scale to more than 2 GPUs, RTX 3090s will be a pain. Each of them takes up 3 slots and consume too much power (~350W). So you have to account for a large case, probably a new motherboard with enough and spacious PCIe slots, and a power supply with the highest power delivery you can find.

Depending on the prices in your region, I suggest you take a look at RTX A5000 too. In my region, they sell for a little lower than RTX 3090, they take up only 2 slots, and they run with much lower power (~230W). They also have vGPU software support. Their performance is a bit lower than 3090’s (but still amazing), but their pros can’t be ignored.

5 Likes

Hi, you can see this post:

3 Likes