If Bitcoin collapses, we're going to see the market get flooded with cheap graphics cards. ML researchers will be able to build a GPU supercomputer for pennies on the dollar.
Conversation
Replying to
I remember when we could buy brand new 1080TIs for under $500. With the price increase over the last few generations, I wonder if will ever return to that era.
2
1
Also the really useful GPUs for deep learning nowadays are the high VRAM ones, of which only the 3090 is popular for mining. Trying to find that GPU in a 2 slot configuration is critically virtually impossible nowadays before even going into the thermal issues.
1
1
Replying to
In terms of a CUDA device with that much VRAM, it’s in own class. The Ax000 series is good enough for training all but the largest models and what most researchers would want to buy. They start at about $4000 per GPU though for the same VRAM as the 3090.
2
1
There are models that go up to about ~50GB vram, which are the most appealing, but the pricing also scales linear with VRAM.