Old enterprise GPUs can cut the cost of running AI locally
Tests of 15 retired NVIDIA enterprise GPUs found that some remain useful for low-cost modern AI work. A 16GB P100 sells for about $75, while a 16GB V100 can cost less than $200. Paired with cheap X99 Xeon motherboards, these cards offer a large amount of VRAM for a modest upfront price.
The V100 delivered performance close to the much more expensive T40, making it the best overall value in the tests. For , the P40 performed better than the P100 within NVIDIA's older Pascal generation. Software support has ended for many of these cards, but older tools such as llama.cpp can still work when compiled from source.
Their poor power efficiency remains a drawback, though a home setup can be switched off between AI jobs. The test suite covered K80, M10, M40, M60, P40, P100, V100, T40 and other cards across , , Blender, Whisper, and additional workloads.
Key points
- A 16GB P100 costs about $75, and a 16GB V100 can cost under $200.
- The V100 performed close to the more expensive T40 and offered the best balance.
- For , the P40 was a better choice than the P100.
- Older software such as llama.cpp may need to be compiled from source.
- Turning the machine off between AI jobs can limit the power cost of old GPUs.