← sources

I went a little over my AI budget this year...

source: reddit (reddit:homelab)published 2026-10-03 17:25ingested 2026-10-04 02:38

retained text

This is basically the 2026 evolution of my old 64-core SFF EPYC build I posted here a few years ago.

https://www.reddit.com/r/homelab/comments/182i7k3/mini_itx_epyc_64_core_128_thread_sff_build/

Started this AI rabbit hole Bought two ASUS GX10s, then apparently decided two wasn’t enough and bought a third.

The GX10s are awesome for LLMs, especially larger distributed models, but image/video generation made me want real discrete GPUs bandwidth. So I bought an RTX PRO 6000 Max-Q. Then a second one due to rising costs. That second card is what finally forced me to change the old SFF setup.

I almost squeezed everything into a Jonsbo Z20, but dual 96GB GPUs were just too tight. I’ve always hated huge cases, so instead of going bigger I went open-frame.

Current damage:
* EPYC 7662, 64C/128T
* 384GB ECC DDR4
* Two RTX pro 6000 Generation GPUs, 96GB each - one in a PCIe 4.0 slot and a dual SAS connection with 2-port bifurcation to achieve x16 speed.
* ASRock Rack ROMED4ID-2T
* Icy Dock with 2x 2TB NVMe over SlimSAS
* m.2 nvme 4tb
* 3x ASUS GX10/GB10 on a 200Gb ConnectX-7 ring

I’d love to eventually rebuild the main box around SP6, but with what newer CPUs and DDR5 ECC cost right now, this old EPYC platform can keep earning its keep.
Went a little over the AI budget this year. 

At least now I have enough local compute to ask an LLM where all my money went...

document fetched 2026-10-04 00:32 · canonical https://reddit.com/r/homelab/comments/1wwsrnn/i_went_a_little_over_my_ai_budget_this_year