- Learn
- Pick your hardware
The Tesla P40 for local LLMs: 24 GB cheap, but is it fast enough?
What the 24 GB Tesla P40 runs, how fast, how it compares to the P100, and the cooling, power and software catches before you buy one.
The Tesla P40 is a 2016 data-center card with 24 GB of memory, the same amount as an RTX 3090. That's why people buy it. The catch is speed: its memory moves about 346 GB/s, well under half of a 3090 and less than half of its sibling, the Tesla P100. It fits big models; it doesn't run them quickly.
What it runs
Those are our calculator's estimates. The short version:
- Mixture-of-experts models are its sweet spot. Qwen3 30B-A3B and gpt-oss 20B only read a few billion weights per token, so they run at an estimated 60 tok/s or so.
- Small dense models are fine. Qwen3 14B at around 20 tok/s is comfortable for chat.
- Big dense models are slow. Qwen3 32B fits, but at an estimated 10 tok/s. Gemma 3 27B is about the same.
Why the gap? Generating text is limited by how fast the card can read its memory, not by its compute. See why bandwidth sets speed and mixture of experts explained.
P40 or P100?
| Tesla P40 | Tesla P100 | |
|---|---|---|
| Memory | 24 GB GDDR5 | 16 GB HBM2 |
| Bandwidth | 346 GB/s | 732 GB/s |
| Qwen3 14B, estimated | ~20 tok/s | ~43 tok/s |
| Qwen3 30B-A3B on one card | Fits | Doesn't fit |
The P40 wins on capacity per card: one P40 holds a 30B-class model, while you need two P100s. The P100 is roughly twice as fast at everything that fits. We chose P100s for our rack, and two of them run Qwen3.6-35B-A3B at 53 tok/s, measured. Full story in the Tesla P100 guide.
If you want a single passive card for a 30B MoE model and speed isn't critical, the P40 makes sense. If you have the slots for two cards, two P100s are usually the better buy.
Two P40s
Two P40s give you 48 GB, enough for Llama 3.3 70B at 4-bit. Our calculator puts that at about 4 tok/s, which is readable but slow. For 70B at a comfortable speed, two RTX 3090s are the standard answer, at about 12 tok/s estimated.
The catches
It has no fans. Like the P100, the P40 is passive and expects server airflow. In a desktop case it overheats quickly. Fit a blower shroud or run it in a server chassis. See cooling passive data-center cards.
It uses a CPU-style power plug. The P40 takes an 8-pin EPS connector, not a standard graphics card plug. Use a proper adapter cable, and never force a PCIe cable into it.
Weak half-precision math. Pascal's P40 is very slow at 16-bit floating-point work. llama.cpp and the tools built on it (Ollama, LM Studio) handle this well, which is why they're the right choice here. Engines that rely on fast 16-bit math run poorly or not at all.
Old architecture. NVIDIA's newest CUDA releases have dropped Pascal. Stay on a driver and CUDA version that still supports it, and don't update blindly.
Slow at long prompts. No tensor cores means reading a long document takes a while, even when generating the reply is fine.
No display outputs. You need another GPU or the CPU's graphics for a monitor.
What to pay
Our builds page doesn't track P40 prices, so check current used prices. As a rule, compare the price against a P100 (we paid about $85 each) and a used RTX 3090 ($650 to $850 in October 2026). The P40 only makes sense well below the 3090.
Who it's for
- Good fit: a homelab box that needs 24 GB on one card cheaply, and mostly runs MoE models like Qwen3 30B-A3B.
- Bad fit: dense 32B models at speed, image generation, or a quiet desk machine. A used RTX 3090 does all three.
Full speeds are on the Tesla P40 page. If you run one, add your numbers to the leaderboard.
One email when there's a new guide, podcast episode, or result from the rack. No spam.