TL;DR: A used Tesla P40 sells for $239–$330 in August 2026 — the cheapest 24GB of VRAM on the market by roughly $500. It decodes 7B models at a measured 54.7 tok/s and holds model files nothing under $800 can hold. The price is low because the card demands three mods before it boots and NVIDIA has already scheduled its driver death: the 580 branch is Pascal’s last, supported only into mid-2028.

	Used Tesla P40 (~$260)	Used RTX 3060 12GB (~$286)	Used RX 7900 XTX (~$815)
Best for	Most VRAM per dollar, period	Plug-and-play budget 12GB	Cheapest modern 24GB
VRAM / bandwidth	24GB / 347 GB/s	12GB / 360 GB/s	24GB / 960 GB/s
The catch	Cooling mod, EPS power, drivers end mid-2028	Half the VRAM	~3× the price, no CUDA

Honest take: Buy the P40 only if you enjoy the project and your workload is llama.cpp/Ollama inference on Linux — as a tinkerer’s second card or a headless inference box, $260 for 24GB is still unmatched. If you want a card that just works, the same money buys a RTX 3060 12GB with no asterisks, and if you want 24GB you’ll actually use in 2029, save for the 7900 XTX or a 3090.

Four days ago we called the used RX 7900 XTX the cheapest 24GB card of August 2026 at $815. That’s true if you mean a card you can game on, plug a monitor into, and expect drivers for in 2030. Strip away all three expectations and the floor drops by more than $500: the NVIDIA Tesla P40, a 2016 datacenter inference card, lists used from $239 and new-old-stock from $329 on GPUDojo’s August 2026 tracker, with typical shipped eBay listings landing in the $260–$330 band.

That is the same money as a used RTX 3060 12GB — $286 on ResalePrices — for twice the VRAM. In a market where the DRAM crisis holds the used RTX 3090 at $1,264 almost entirely on the strength of its 24GB, a $260 card with the same capacity sounds like a pricing error.

It isn’t. The market knows exactly what the P40 is: the best VRAM deal in local AI, wearing three project-card asterisks and an expiration date. Here’s the honest math on both.

The numbers that matter

The P40’s spec sheet, from NVIDIA’s official product brief and Microway’s technical breakdown: Pascal GP102 silicon (the RTX-less ancestor of the GTX 1080 Ti), 3,840 CUDA cores, 24GB of GDDR5 on a 384-bit bus, 347.1 GB/s of memory bandwidth, 250W board power, launched September 2016 as a $5,699 datacenter inference card. Passive cooling, no display outputs, and a CPU-style 8-pin power socket — more on all three below.

Two spec-sheet details decide everything about how this card behaves in 2026:

The bandwidth is 3060-class. Token generation is memory-bound — the GPU re-reads every active weight per token — so 347 GB/s means the P40 decodes like a RTX 3060 12GB (360 GB/s), not like the 24GB cards it matches on capacity. The 3090 moves 936 GB/s, 2.7× more. The P40’s 24GB doesn’t buy speed; it buys admission — models the 3060 can’t load at all.

There’s no fast FP16. GP102 runs half-precision at 1/64th the FP32 rate — Microway measured the P40’s FP16 as effectively unusable; the card was built for INT8 inference (47 TOPS). This kills PyTorch-based diffusion and fine-tuning workloads, which assume FP16/BF16. llama.cpp is the exception that makes the card viable: its quantized kernels route around the FP16 hole, which is why every working P40 rig you’ll find runs GGUF models.

What that delivers, measured: on llama.cpp’s official CUDA scoreboard (commit c76b420, August 2025), the P40 runs Llama 2 7B Q4_0 at 1,007 tok/s prompt processing and 54.7 tok/s generation — with flash attention enabled, 1,079 and 53.7. On current-generation models, SarmaLinux’s Ollama-on-P40 playbook measures Gemma 3 12B Q4_K_M at 25.3 tok/s. Both land where the 347 GB/s math says they should.

Here’s the used market sorted by what a dollar buys at each card’s actual job (prices: GPUDojo and ResalePrices, August 2026; ratios are our arithmetic):

Card	Typical used price	VRAM	Bandwidth	$ per GB of VRAM
Tesla P40	~$260	24GB	347 GB/s	~$11
RTX 3060 12GB	~$286	12GB	360 GB/s	~$24
RX 7900 XTX	~$815	24GB	960 GB/s	~$34
RTX 3090	~$1,264	24GB	936 GB/s	~$53

Eleven dollars per gigabyte. Nothing else on that table is under twenty. That’s the entire case for the card — now the case against it.

What 24GB at 347 GB/s actually runs

Check your exact model-plus-context combination in our VRAM calculator first; the numbers below are the shape of the tier.

The sweet spot is MoE models, and it’s a genuinely good spot in 2026. Sparse mixture-of-experts models read only their active parameters per token, so decode speed scales with active size while VRAM demand scales with total size — the exact inversion of the P40’s strengths and weaknesses. A 30B-class MoE with ~3B active parameters fits its ~17GB Q4 file entirely in the P40’s 24GB and re-reads only ~2GB per token. This is why the card aged better than its bandwidth suggests: in a February 2026 llama.cpp thread, a builder running Tesla P40s reports “over 20 TPS on most smaller MoE models” — and 5.5 tok/s on the 1T-parameter Kimi K2.5 across eight of them. The best-models-for-24GB guide is effectively this card’s software manual, with one adjustment: weight its MoE picks and ignore its dense-27B speed numbers.

Dense 13B and under runs acceptably. The measured 25.3 tok/s on Gemma 3 12B is comfortably past reading speed, and the 24GB means you run it with 32K context instead of the 8K a 12GB card allows.

Dense 27B+ fits but crawls. A 27B Q4_K_M file is ~16.5GB — loads with room to spare, but 347 GB/s over 16.5GB of weights puts the theoretical ceiling at ~21 tok/s and the realistic number at 11–14 tok/s — a bandwidth-scaled estimate, not a measurement, extrapolated from the measured 12B result. Usable for patient chat; frustrating for agentic coding workflows, where long generations amplify every token of latency.

Diffusion and fine-tuning are off the table. No fast FP16 means Stable Diffusion and Flux run somewhere between badly and not at all, and QLoRA fine-tuning — even where VRAM suffices — assumes compute this card doesn’t have. If that’s your workload, this is the wrong article; start with the GPU buying guide.

The three mods before it boots

The P40 was built to slide into a server chassis that provides airflow, power, and a management plane. On your desk, all three are your problem — this is where the $500 discount goes.

Problem one: it has a CPU power socket, not a PCIe one. The 8-pin connector on the card’s edge is EPS-keyed (like a motherboard CPU connector), not PCIe-keyed — NVIDIA’s product brief specifies the CPU 8-pin, and plugging a standard PCIe cable in anyway is a documented way to kill the card. The fix is a dual PCIe 8-pin to EPS adapter cable (NVIDIA part style 030-0571-000, sold by COMeap and others) — Level1Techs maintains a sourcing thread because counterfeit pinouts exist. Buy from a seller with the pinout photographed.

Problem two: it ships with no fan. The card is a passive brick expecting ~60 CFM of forced server airflow. Without it, the P40 hits thermal limits in minutes: one measured desktop build found the card throttling even after power-limiting to 130W, and generation speed jumping from 3.6 tok/s to a consistent 10 tok/s once a sealed fan duct was fitted — the same silicon, nearly 3× the speed, purely from cooling. The community fix is a 3D-printed shroud with a radial blower — free STLs on Printables — or a pre-made kit from GPUmods or eBay. Budget for the blower to be audible; high-static-pressure fans are.

While you’re at it, power-limit the card. The last 100W buys very little inference speed on a bandwidth-bound workload:

$ sudo nvidia-smi -pl 140
Power limit for GPU 00000000:01:00.0 was set to 140.00 W from 250.00 W.
All done.

At the stock 250W, an hour of load costs $0.030 at $0.12/kWh (our arithmetic — your rate applies); at 140W it’s $0.017, and the blower works far less. The 24/7 home server power math applies double to a card this thirsty per token.

Problem three: no display outputs. The P40 is compute-only. You need integrated graphics or a second GPU for your monitor, and on consumer motherboards you may need to enable Above 4G Decoding in UEFI before the card enumerates. Plan on it living headless in a Linux box — which is where its remaining driver support is longest anyway.

The clock: mid-2028, and what actually stops working

This is the part the eBay listings don’t mention. NVIDIA has formally ended Pascal’s road: CUDA 13.0 dropped compilation support for everything below compute capability 7.5 — Maxwell, Pascal, and Volta — and the 580 driver branch is the last that supports these cards, maintained as a long-term support branch only until mid-2028. A P40 box pins itself to CUDA 12.9 and driver 580.x, permanently:

$ nvidia-smi --query-gpu=name,driver_version --format=csv,noheader
Tesla P40, 580.65.06

What this means in practice, honestly assessed: inference keeps working long past 2028 — a frozen driver doesn’t stop functioning, and llama.cpp still ships CUDA 12 builds today. The Pascal community is active enough that the February 2026 llama.cpp thread is largely P40 owners posting configs (LLAMA_CUDA_FORCE_DMMV=1 remains the recurring tuning flag) and the maintainers haven’t dropped compute 6.1. But the direction is one-way: each new PyTorch, each new CUDA-13-only kernel, each new model architecture optimized for tensor cores widens the gap. You’re not buying a card with 24GB of VRAM; you’re buying a fixed appliance that runs 2026’s GGUF stack at 2026 speeds until its capacitors quit. Price that in.

Buy or skip

Buy it if you’re the target user, which is narrower than the price suggests: you run llama.cpp or Ollama on a headless Linux box, your models of choice are quantized MoEs or ≤13B dense, you own or can print a fan shroud, and you’d rather spend $300 and a weekend than $1,264. As a second card dedicated to holding a model while your main GPU does everything else, it’s arguably the best $260 in the hobby.

Skip it if you want one card that does everything — the RTX 3060 12GB at $286 has warranty-era drivers, a fan, a display output, and runs the 14B class that covers most people’s daily use. Skip it too if your 24GB ambition is real: the used RX 7900 XTX at $815 decodes ~3× faster with a 2030s future, and the used RTX 3090 at $1,264 adds full CUDA on top. And if what you actually need is occasional big-model access, skip the hardware entirely — an hour of rented cloud GPU on RunPod costs less than the P40’s fan shroud and comes with 48–80GB attached.

The P40 in 2026 is the mechanical watch of local AI: outdated on every spec, kept alive by enthusiasts, and — for exactly the person who enjoys that — worth every penny of its very small price.

FAQ

Can the Tesla P40 run a 70B model? It can load one only in pieces — a 70B Q4_K_M is ~42GB, so a single P40 offloads barely half and decode drops to CPU-bound single digits. Pairs and quads of P40s are a classic budget path to 48–96GB (the February 2026 llama.cpp thread runs eight), but at 347 GB/s per card, expect low single-digit tok/s on dense 70B — the 48GB tier guide explains why dual 3090s remain the sane version of that plan.

Does the P40 work with Ollama? Yes, today — Ollama bundles llama.cpp CUDA builds that still support compute 6.1, and SarmaLinux’s playbook documents a working 2026 setup. The risk is forward-looking: once Ollama moves its bundled runtime to CUDA-13-only builds, P40 owners will need to pin old versions or build llama.cpp themselves against CUDA 12.9.

Is the P40 better than the older Tesla M40 or the P100? Than the M40 (Maxwell, 24GB, 288 GB/s): yes on every axis, and the M40’s driver clock is the same — don’t save the $60. The P100 is the opposite trade: 16GB of HBM2 at 732 GB/s with working FP16 — faster per token, but 8GB less admission, which for LLM work is usually the wrong trade.

Why is a card with 24GB this cheap when VRAM prices are exploding? Because datacenters retired P40s by the pallet as inference moved to Ampere and Hopper, and the buyers who set GPU prices — gamers and diffusion users — can’t use it: no display outputs, no fast FP16, no official Windows gaming driver path. The supply is huge and the demand is one hobby. That mispricing is the opportunity, and the mid-2028 driver deadline is why it never fully closes.

Sources

• Tesla P40 24GB Used Price & History, August 2026 — GPUDojo
• Tesla P40 24GB live listings — eBay
• Tesla P40 GPU Accelerator Product Brief (PB-08338-001) — NVIDIA
• NVIDIA Tesla P40 (Pascal GP102) Up Close — Microway
• Performance of llama.cpp on NVIDIA CUDA (official scoreboard, P40 rows) — ggml-org/llama.cpp Discussion #15013
• Running older Pascal GPUs with llama.cpp, February 2026 — ggml-org/llama.cpp Discussion #19248
• Navigating GPU Architecture Support: A Guide for NVIDIA CUDA Developers — NVIDIA Technical Blog
• NVIDIA Confirms 580 Linux Driver Is the Last for Maxwell/Pascal/Volta — Phoronix
• NVIDIA to axe Maxwell, Pascal, and Volta with end of driver support — Tom’s Hardware
• Cooling a Tesla P40 in the living room (measured throttling and duct results) — ariezlabs
• NVIDIA Tesla P40 radial fan shroud, free STL — Printables
• Running Ollama on a Tesla P40, Playbook — SarmaLinux
• Sourcing EPS to dual 8-pin PCIe power adapters for Tesla P40s — Level1Techs Forums
• Used RTX 3090 price — ResalePrices
• Used RTX 3060 12GB price — ResalePrices

Last updated August 30, 2026. Prices and specs change; verify current rates before purchasing.

Recommended Gear

• NVIDIA Tesla P40 24GB — the card itself; compare against eBay before buying
• Tesla P40 EPS power adapter cable (030-0571-000 style) — mandatory; the card’s socket is CPU-keyed, not PCIe
• RTX 3060 12GB — the same-price plug-and-play alternative