NVIDIA Tesla P40
NVIDIA · Budget
A 2016 Pascal data-centre accelerator with 24GB of memory, repurposed as a cheap high-VRAM card for homelab local inference.
Specifications
| VRAM | 24GB |
|---|---|
| Memory bandwidth | 346 GB/s |
| Power draw | 250W |
| Type | accelerator |
A cheap used homelab card, though cooling and power-adapter costs add up. Verify current pricing.
NVIDIA Tesla P40: common questions
- What models can the NVIDIA Tesla P40 run?
- With 24GB of VRAM it can run models up to roughly 37B parameters at a 4-bit quantisation, or smaller models with more context. Use the hardware matrix for specifics; these figures are approximate.
- How much power does the NVIDIA Tesla P40 draw?
- About 250W under load, so pair it with a power supply that has real headroom.
Strengths
- 24GB at a very low used price, the classic budget entry to larger local models
- Widely supported in llama.cpp and other inference stacks
- A passive 250W design suits a server chassis with existing airflow
Weaknesses
- Only 346 GB/s bandwidth, so generation is slow versus modern cards
- Weak FP16 and no tensor cores, so it suits quantised GGUF workloads
- The passive cooler needs added fans in a desktop, and it uses a server power connector
Entry last verified 18 August 2026. Specifications and especially prices change; verify before buying.