If the cryptocurrency market completely crashed, they would be handing out top-end graphics cards by the dozen to gamers for like $99.99.
Conversation
Replying to
As a ML researcher, I wish that were the case. We would be buying them all up way before they got anywhere near that price.
1
2
Replying to
Hey Aaron -- quick question. Do you think the 3060 Nvidia card with 12GB ram is a decent card for smaller training sets? Have you used a 3050/60/70/80/90 before?
1
Replying to
Yeah, it's specs are pretty similar stats to a 1080TI, which are the most common. 3090 is a absolutely great price point at MSRP for a 24GB card, but they are hard to use in multi-GPU setups. The dataset size doesn't affect the VRAM requirements, model size does.
1
1
I've mostly used 3090s as the VRAM on those cards make it the most attractive by far of the current generation. If you can find the discontinued blower fan models, even better as you could fit 3 or even 4 in a single 4 slot motherboard without using PCI risers.
1
1
To clarify, the 3060 specs are pretty similar to 1080TI, (like 30% faster). Any 12GB+ VRAM card will work reasonably. 24GB+ VRAM is even better. I keep hearing rumors of Nvidia dropping a 3000 series 16GB card, since the memory bus supports it, but I have yet to see one listed.
3
1
Replying to
Thanks man! You are awesome! I'd love to talk to you some time voice once we set up our ML servers to get your advice on some things. Happy to pay for your time as well and/or give you time on the large servers.