$360 for 160GB VRAM: building a local LLM server on unlocked CMP 50HX mining cards
Boricua-vet · reddit · 2026-09-25
The author upgraded a local LLM server with four CMP 50HX mining cards ($80 each, $360 total, 160GB VRAM) replacing P102-100s, getting good speeds on local models at just 8W idle. Key details: cards must be inserted perfectly straight in the narrow P520 chassis (a scrape cost them a $50 motherboard repair), the cards need the open-source cmp50hx-unlock tool to be usable, and fine-tuning via RunPod costs only $3-5 per model ($50/year for 10 models), making expensive GPUs unnecessary. The old P102-100 server now serves fine-tuned models for their agents and workflows.
More from coding & agent
- tool-journal: SQLite-based execution journal makes agent tool-call crashes deterministically recoverable — JeremyCMorgan · 2026-09-25
- AutoGPT in practice: automation just promoted me to the worst middle manager ever — Full_Collar9026 · 2026-09-25
- How to pick a model per step in a multi-step agent? Devs weigh routing and evals — LeviYagami · 2026-09-25
- NicheDB launches a dedicated MCP servers directory for the community — Minimum_Hour519 · 2026-09-25
- Coding models now make niche software nearly free to build, down to the details — nabla_theta · 2026-09-25
- Skill-Inject Benchmark Shows Frontier Agents Fall for Malicious Skills, Accepted at NeurIPS 2026 — maksym_andr · 2026-09-25