logoalt Hacker News

chao-today at 3:37 PM1 replyview on HN

Maybe not insurmountable in the "laws of physics" sense?

However, it may be so costly in cobbled-together parts, and in time to deal with unsupported drivers and/or VBIOS, that it is not worth it compared to using a proper server hardware with proper SXM sockets. Nvidia is also doing a lot of new, exciting things with networking that may make their SXM-based GPUs require even more just-so hardware support as time goes on.

Getting older SXM3 GPUs (i.e. for V100's, from 2017) working via PCIe adapters has been done reliably. However, here is someone who did that successfully, and spent a chunk of time last year trying to do the same for SMX5 (H100) and failing:

https://forums.servethehome.com/index.php?threads/sxm5-h100-...

We were all spoiled by the era from 2005-2020, when you could squint at a "server" configuration and see that it was expensive, high-binned, commodity hardware with some extra RAS and OOBM features. You could buy parts harvested from a retired server based on Xeon E5-1680v3 and drop them in your workstation. Or you could buy an entire single-socket Xeon E5 v3 server, plop in a GPU, and mostly use it as a workstation! As-is!

Now the serious servers have SOCAMM memory, 400Gbps networking, EDSFF SSDs, maybe are configured for CXL 3.0, et cetera. The hardware itself is so divergent that it isn't swap-in-plug-and-play with desktops, even for some high-end workstations.


Replies

nixon_why69today at 4:28 PM

I really appreciate your comment and the difficulties, but if we're hypothetically talking about $100k hardware for pennies on the dollar.. it looks like the problems are all analog? Like ok, add power supply on top of form factor and cooling but all of the silicon is compatible, right? If (if!) we are talking about pennies on the dollar then those problems are solvable.

For example, if I am a homelab, I don't necessarily need the integrated SFP networking stuff to work, I'm happy with my single overpowered GPU. I don't need CXL either, I just want one badass H200 running in my rig. Maybe Shenzhen will productive that?