Local AI Hardware / GPUs for local AI

NVIDIA Tesla P40 24GB for local AI

Sold prices for the NVIDIA Tesla P40 24GB are tracked every day on GPU Sold (https://gpusold.com/gpu/tesla-p40).

Cheap 24 GB, but old, slow, and passively cooled.

Memory
24 GB
Memory bandwidth
347 GB/s
Board power
250 W
Architecture
Pascal

What matters for AI

The P40 has no fan; it needs a server chassis or a fan shroud. Pascal also lacks fast FP16, so it is slow at prompt processing compared with anything newer.

Before you buy a used GPU for AI

  1. Memory first. The model and its context have to fit in VRAM (or it spills to system RAM and slows to a crawl). Use the VRAM calculator.
  2. Bandwidth second. Token generation speed is mostly limited by memory bandwidth, not compute.
  3. Check memory temperatures under load on GDDR6X cards (3090, 3090 Ti). Worn thermal pads are common on ex-mining cards.
  4. Plan power and slot spacing before buying two or more cards. Transient spikes on high-end cards trip undersized power supplies.