DIY CUDA box with unlocked CMP170HX mining cards hits 4000 tps prompt processing for Qwen Flash Next
Miserable-Dare5090 · reddit · 2026-09-02
A Reddit user shared a full build log for a DIY CUDA server running LLMs on two unlocked CMP170HX mining cards.
- Expanded PCIe lanes via cheap Newegg CPU/MB combos; reused 64GB DDR5 and SSDs from an old prebuilt
- Budget case + Noctua fans; both CMP170s in a GPU switch on the main x16 slot, other cards on oculink
- Qwen Flash Next fits nicely: 4000 tps prompt processing, 80+ tps single-stream decode without MTP
- DeepSeek repo would need at least 3x 64GB cards — risky as mining card prices rise
- Now runs three models simultaneously: Flash Next, Qwen 27B, Gemma 26B
More from Infra
- SGLang Summit sponsors grow from 11 to 36 in one week; Lip-Bu Tan, Lilian Weng to speak — BanghuaZ · 2026-10-18
- Yacine jokes his datacenter KPI is gigawatt-hours: get paged if not burning enough power — yacineMTB · 2026-10-18
- 1.68x inference speedup drops, directly accelerating GRPO post-training — yacineMTB · 2026-10-18
- OpenRouter is only a fraction of the total addressable market, poster argues — sloppenheimer · 2026-10-18
- Developer builds change alerts for LLM prices and retirements across Bedrock, Azure and Vertex — jsandhu0780 · 2026-10-18
- 12.3M tokens in 27 minutes for $0.14: agent session with 99.6% cache hit — teortaxesTex · 2026-10-18