NVIDIA Tesla P40 review

You’ve gone looking for 24GB of VRAM on a tight budget, and the used Tesla P40 listings keep turning up. It’s NVIDIA’s Pascal datacentre card from 2016, with 24GB of GDDR5 on board. It was made for a server though, so it arrives with no fan, no display output and a CPU power socket where you’d expect a PCIe one, and sorting those out is down to you.

The take

If you’re weighing a P40 against a used RTX 3090, the two hold the same 24GB, and that memory is what the older card has going for it. The P40 has 3,840 CUDA cores on the GP102 chip and up to 347 GB/s of memory bandwidth. It’s a dual-slot, full-height board that draws 250W. On the llama.cpp CUDA scoreboard, a tester posting as m18coppola ran both cards on a 7B model and the 3090 generated text roughly 2.9 times as fast. InsiderLLM’s September 2026 eBay sold-listing pull put a bare P40 at $220-290, and Serverdeals’ asking prices on 12 November 2026 started at about €173, with a median of about €289.

The first job is fitting a fan, since the P40 is passively cooled and hasn’t got one, which means it relies on air pushed through it to stay within its thermal limits. The fix owners describe is a 3D-printed blower shroud with a radial fan on the end, which RBA Consulting reports works extremely well. Power comes through one EPS 8-pin, the CPU type rather than a PCIe plug, though NVIDIA lists a CPU-to-PCIe dongle that takes two PCIe cables. There’s no display output either, so you need another GPU or an iGPU for the screen. It sits in an ordinary PCIe 3.0 x16 slot, but you’ll need Above-4G decoding switched on in the BIOS. The card asks for a 32GB memory window, and Serverdeals notes it isn’t detected without that setting.

Buy a P40 now and you’re buying into a driver line with an end point. NVIDIA has said the release 580 driver series will be the last to support Pascal, and CUDA Toolkit 13.0 removed offline compilation and library support for it. The card doesn’t stop working, it just stays on driver branch 580 and on software built against CUDA 12.x, so anything that needs CUDA 13 won’t run on it. vLLM already doesn’t, because it asks for compute capability 7.5 or higher and the P40 is a 6.1 card.

What it’ll run

On a P40 you’ll be running quantised GGUF models in llama.cpp, because the card is very slow at FP16 (half-precision maths). Its FP16 rate is about 1/64 of the FP32 rate, which works out to about 0.18 TFLOPS against 12. Local AI Master calls FP16 on this card effectively unusable. The August 2025 scoreboard run by m18coppola has a single P40 at about 55 tok/s on Llama 2 7B Q4_0, where the same tester’s RTX 3090 managed 158. Flash Attention works on the P40 as well, thanks to a llama.cpp change merged in May 2024 that enabled it on cards without fast FP16.

The bigger models are what the 24GB is for. A 32B model at Q4_K_M is a 19.9GB file, which leaves about 4GB for context. A 70B at Q4_K_M is 42.5GB, so that one needs two P40s. RBA Consulting reported Qwen3 Coder 30B at roughly 50 tok/s under Ollama in December 2025. That’s a mixture-of-experts (MoE) model, though, with only about 3.3B of its 30.5B parameters active for each token, so it isn’t a guide to how a conventional 32B will run.

Who should buy it, and who shouldn’t

The P40 makes sense if the thing you’re short of is VRAM and the budget is small. It suits someone who’ll print or buy a blower shroud and fit a radial fan, who has a second GPU or an iGPU to drive the screen, and who’s happy running quantised GGUF models in llama.cpp with Flash Attention on. If that kind of faff is part of the appeal, the P40 is a reasonable buy.

It doesn’t make sense if you want to plug a card in and go, or if you need vLLM or anything built on CUDA 13. And if speed matters more to you than memory, a used RTX 3090 has the same 24GB and is the card to price up instead.