LGA1700 to a threadripper?
retained text
Please check my logic, I'm new to counting PCIe lanes :) I tried vllm radiance with 2xr9700 and learned that my (LGA1700 cheap and old) motherboard doesn't support P2P or PCIe atomics. The plan is to build for a fast smaller model in 27b range (I have a slower strix for medium sized models): i) support my 2 GPUs today, with a mobo allows me to \\\*dream\\\* of 4 total GPU in future without 1 card becoming a bottleneck in parallelism ii) focus on vram today, with \\\*possibility\\\* of buying 2nd hand DDR5 RAM \\\*if\\\* DDR6 floods the market in 1-3 years \- Motherboard DDR4 vs DDR5 Going for octa-channel DDR4 would be ok for today, however I'm not building for large moe today so pointless to get locked in Going ddr5 means smaller capacity today (16/32GB for just OS) \- Planning PCIe lanes for 4 GPU Having 3xGen5 slots + 1xGen4 seems like a bottleneck for TP=4 (or future MOE), this rules out Intel Socket 1851 (not enough total PCIe lanes) 4xGen5x16 rules out some TRX50 motherboards. Only a top end combo like 'Gigabyte TRX50 AI TOP-2B' paried with a 7000-WX, but now we're talking over 5k for CPU And mobo. WRX90 can do this, CPU and mobo 2500, but requires an EEB motherboard, EEB case, dual PSU, etc \- Verdict Pure 4xGen5: I can't justify cost/space/usecase (full spec trx50/wrx90) Gen5+Gen4: Good middleground (cheaper trx50), but has future bottleneck, clever bios/bipurification can smooth the difference 2 Gen5x16: stay 'humble' (Socket 1851/z890) Are there any paths or architectures I've missed? I'm looking past Epyc, it's too server like and seems better suited for CPU inference? Many thanks!
document fetched 2026-09-22 00:41 · canonical https://reddit.com/r/LocalLLaMA/comments/1wm95cv/lga1700_to_a_threadripper