OverviewExploreTrending
Nostr Archives
OverviewExploreTrending
ChipTuner14h ago
Did some napkin math last night. You can pretty much buy the exact GPU you rent from a provider in about a year at current prices, including 24/7 electric. That's nuts to me. At least it used to be reasonable in a 60 month hardware lease period. Specifically enterprise GPUs - Id still recommend buying GPUs over renting them. Otherwise buy your inference as usual. https://gpucost.org/
💬 6 replies

Replies (6)

semisol7h ago
Unfortunately I do not have a datacenter to put them in. Also need to consider the server HW I guess, which would probably make it 18 mo instead.
0000 sats
ChipTuner6h ago
Yeah If you need a machine to stick them in that's gonna raise the price a bit right now. The xeon scaleable prices really shot up not including DDR4. Crazy to spend like $2500 on 128gb 2866 and a pair of casecade lake golds.
0000 sats
semisol7h ago
Unfortunately I do not have a datacenter to put them in. Also need to consider the server HW I guess, which would probably make it 18 mo instead.
0000 sats
ChipTuner6h ago
Sats or GPU right now?
0000 sats
semisol5h ago
I meant physical hosting part of things :) Also, I'd maybe wait 1-2 years. We are getting close to the plateau of HW-level innovation instead of standard process scaling but I do not think we are not there yet. LLMs meant we got FP4/8 formats for example, so anyone with pre-Hopper hardware is toast for higher-density models.
0000 sats
ChipTuner4h ago
Good points!
0000 sats
ChipTuner6h ago
Yeah If you need a machine to stick them in that's gonna raise the price a bit right now. The xeon scaleable prices really shot up not including DDR4. Crazy to spend like $2500 on 128gb 2866 and a pair of casecade lake golds.
0000 sats
ChipTuner6h ago
Sats or GPU right now?
0000 sats
semisol5h ago
I meant physical hosting part of things :) Also, I'd maybe wait 1-2 years. We are getting close to the plateau of HW-level innovation instead of standard process scaling but I do not think we are not there yet. LLMs meant we got FP4/8 formats for example, so anyone with pre-Hopper hardware is toast for higher-density models.
0000 sats
ChipTuner4h ago
Good points!
0000 sats