Old server machine with 2x 3060 12GB for MoE models, anything I’m missing?
retained text
I'm about to buy a used HP Z640 for around €550. It has 2x E5-2680 v4, 64GB ECC at 2400, and a Quadro M6000 24GB that comes with it. The seller’s MemTest86 ran 6 hours with no errors. The plan is to add two 3060s in the x16 slots and run Qwen 3.8 Flash Next on Strata, as well as of course normal llama.cpp. I also do PyTorch CV training, RL in PyBullet and plan to do LoRA training for LLMs + video generation with MiniMax H3. Of course the first thing will be installing Ubuntu. There are a few things I’d love to hear from anyone who’s run something similar: * Would it be a problem to fit two 3060s in a Z640? To me it doesn't seem so but I've never actually built a PC, so I doubt myself. * Is dual socket a pain for llama.cpp and MoE offload? Any hidden issues this would cause? * Is there any use for keeping the M6000 as a third card on a CUDA 12 build, or is it dead weight? (I know Strata won’t use it, I mean for llama.cpp) * CPU hit 86°C in the seller's MemTest run. Normal for this chassis? * I read the PSU only has two 6-pin cables and HP limits two cards to 150W each with two CPUs, so I plan to use adapters and power limit the 3060s. Is this the right approach? In general I am super thankful for any input, so please correct me about anything!
document fetched 2026-10-21 00:41 · canonical https://reddit.com/r/LocalLLaMA/comments/1xamrvy/old_server_machine_with_2x_3060_12gb_for_moe