Building a 4-GPU local LLM rig: why Threadripper beats LGA1700 on PCIe lanes
El_90 · reddit · 2026-09-21
After discovering his LGA1700 motherboard lacks P2P and PCIe atomics support when running vLLM with 2x R9700, a Reddit user works through upgrade paths for a rig targeting 27B-class models today and 4 GPUs later.
Key reasoning: octa-channel DDR4 only matters for large MoE models; DDR5 offers less capacity now but better future headroom. On PCIe, 3xGen5 + 1xGen4 bottlenecks TP=4, ruling out Intel Socket 1851; full 4xGen5x16 requires either a $5k+ TRX50 combo (Gigabyte AI TOP-2B + 7000-WX) or $2.5k WRX90 with EEB case and dual PSU.
Verdict: pure 4xGen5 isn't justifiable; Gen5+Gen4 is a middle ground with future bottleneck; or stay humble with 2xGen5x16 on Z890. He asks the community for missed architectures, skipping Epyc as too server-oriented for his use case.
More from Infra
- Rothschild Redburn goes bearish on AI compute, rates NBIS and CRWV Sell — JOBhakdi · 2026-09-21
- PC memory module prices keep climbing daily as Meta's Muse holds #1 on the App Store — firstadopter · 2026-09-21
- Quantizing Cellpose-SAM for stem cell imaging: W4/W8 hits 6.76x compression with zero failures — capicu-ai · 2026-09-21
- Report: DeepSeek training a 2T-param model with 8T planned, Huawei chips due late 2026 — Hesamation · 2026-09-21
- CXMT's wafer capacity to grow 10x by 2030, nearing Micron parity in late 2026 — Terminator857 · 2026-09-21
- Komlós conjecture solution announced, with overlooked implications for neural network quantization — stevenstrogatz · 2026-09-21