Post #3845007
2026-07-15 22:59 UTC
@azonenberg@ioc.exchange @Lydie@tech.lgbt This.
Basically those "GPUs" if we can still call them that, are gigantic monoliths. It's wasted silicon.
The best thing we could do is possibly make tiny boards that can with some custom firmware successfully boot one or two of those monoliths. Buy then for what use?
The sleds are insanely dense and power consuming, so just turning them on is a no-go most of the time.
Really don't see a bright future for anything DC after the AI bubble. I hope I am wrong, but I sincerely don't know how.
Replies (1)
-
@azonenberg@ioc.exchange 2026-07-15 23:29
@YvanDaSilva@hachyderm.io @Lydie@tech.lgbt I mean, if I could a) get working Vulkan drivers rather than only CUDA, b) convert the rack-level water cooling loop to a standalone AIO or air cooled heatsink, c) get the NVLink specs opened up to enable third party interoperability, I would definitely find a use for one single MGX sled in my lab. I'd probably unplug one of the two processor boards inside to save power as one would likely be enough, but I have enough heavy number crunching I'd put it to use. But reworking all of the hardware and software to enable that would likely be more work than just getting workstation/not-insane server parts.