$100 used Tesla P100 beats gaming GPU nearly 2x for local LLM inference
Mrinohk · reddit · 2026-09-28
A local inference enthusiast shared a hands-on test that contradicted Claude's advice against used Tesla P100s (no tensor cores, weak quantization, power/cooling hassle). After buying one for under $100 and applying community llama.cpp patches with --cpu-moe, results far exceeded expectations.
Key numbers: on an RX6600 XT, Qwen3.6 35B A3B (Q4 + MTP) ran at 30-35 tok/s generation (45-50 for code); the P100 with shinbunbun's dedicated patches hit 54-60 tok/s (66-72 for code) at 32k context. Cooling uses a 3D-printed housing with a 94mm Noctua fan; a new PSU cost about $100.
Takeaway: don't fully trust frontier models on budget hardware. The P100 is a seriously underrated entry card for local inference—nothing else comes close at that price. A second card is on the way for a dedicated dual-GPU inference box.
More from Infra
- India's Sarvam launches Token Factory, hosting top open models on local infra — HarveenChadha · 2026-11-16
- Gemini 3.5 Flash now runs on air-gapped Google Distributed Cloud, offline up to 12 months — shashib · 2026-11-16
- Gemma.c: full Gemma 1/2/3 multimodal inference in a single C file, ~72 tok/s on CPU at INT8 — This-Purchase-3325 · 2026-11-16
- BIS Issues Interim Final Rule Addressing Remote Access to Advanced AI Chips — apples_jimmy · 2026-11-16
- Jailbroken PS4 runs a 9B open-source model at up to 43 tok/s by talking directly to its 2013-era GPU — AIFlow_ML · 2026-11-16
- Mobile AI Factory: agents auto-build on-device SLMs with 19x less GPU time — pmttyji · 2026-11-16