Tesla P40

The Tesla P40 is the 2016 Pascal-architecture datacenter card that became a used-market legend for budget LLM plebs. 24 GB of GDDR5 on a 384-bit bus yields 347 GB/s — roughly a third of a 3090 — but at \$150–250 on the used market it is the best \$/GB of VRAM you can buy. Pascal descends directly from Maxwell (GTX 980 era) and predates Volta’s first-gen tensor cores, so inference is pure FP32/FP16 compute with no tensor acceleration.

Who it’s for: budget-first plebs building their first LLM rig, or anyone stacking VRAM (two P40s = 48 GB for \$400). Acceptable for background/batch workloads where tok/s is not critical.

Models it runs comfortably: Llama 3 8B at Q8, Llama 3 70B distillations at Q4 (slow but functional), Mixtral 8x7B at Q4. Expect 5–10 tok/s on 70B-class — fine for chat, painful for agents.

Hashcenter notes: passive blower — designed for server airflow, needs a fan shroud in a tower case. 250 W TDP on an 8-pin EPS (not PCIe) connector — requires an adapter. No display output, headless only. Single-slot blower form factor makes it excellent for 4-GPU rack builds. Credit NVIDIA for building the card and the secondary market for making it affordable for plebs.

Further reading: This card is a core component of a pleb-grade AI Hashcenter. Pair it with the sovereignty argument in the Sovereign AI for Bitcoiners Manifesto, or look at how the same 120V envelope powers a Bitcoin space heater in our mining catalog. Running both workloads on one rig? See Heating Your Home With Inference.