Skip to content
Llama Builds

NVIDIA · Pascal · 2016

NVIDIA Tesla P40 for local AI

The cheapest 24GB card, if you can cool it and accept Pascal's slow compute.

Updated · Speeds marked ~ are estimates; seemethodology

VRAM
24GB
Bandwidth
346 GB/s
Power
250W
Runs well
8 models

Tesla P40

Typically about $230 used

Check price at AmazonFind it used on eBay

Can the Tesla P40 run LLMs?

The Tesla P40 has 24GB of GDDR5 and 346 GB/s of memory bandwidth, which decides how big a model fits and how fast it generates text. It comfortably runs models up to about 35B parameters: the largest in our catalog is Qwen3.6 35B-A3B at Q3_K_M (19GB with 8K context, ~58 tokens/sec). Our top all-round pick for it is Qwen3.8 27B (Q4_K_M, ~11 tokens/sec). Llama 3.3 70B Instruct only runs with partial CPU offload, at about 1.6 tokens/sec with CPU offload.

Recommendations

Best models for the Tesla P40

  1. Best overall · also coding, reasoning, vision

    Qwen3.8 27B

    Q4_K_M · 18GB · ~11 tokens/sec

    Runs well

  2. Runner-up

    Gemma 4 26B A4B

    Q5_K_M · 20GB · ~36 tokens/sec

    Runs well

  3. Fastest good model · also largest that fits

    Qwen3.6 35B-A3B

    Q3_K_M · 19GB · ~58 tokens/sec

    Runs well
Read: the best LLMs to run on the Tesla P40

Compatibility

What runs on the Tesla P40

Every current model in our catalog, at the best quantization that fits in 24GB with 8K tokens of context. "Partial offload" means part of the model runs from system RAM (we assume 64GB), which is much slower.

Which open-weight models run on the NVIDIA Tesla P40, with the best quantization and estimated speed
ModelParametersBest quantMemorySpeedVerdict
Gemma 4 E4B8BQ8_010 GB~20 tok/sRuns well
Qwen3.5 9B9BQ8_012 GB~17 tok/sRuns well
Phi-414.7BQ8_018 GB~12 tok/sRuns well
gpt-oss-20b20.9B (3.6B active)MXFP415 GB~41 tok/sRuns well
Gemma 4 26B A4B26B (3.8B active)Q5_K_M20 GB~36 tok/sRuns well
Qwen3.8 27B27BQ4_K_M18 GB~11 tok/sRuns well
Gemma 4 31B31BQ4_K_M22 GB~9.6 tok/sFits, tight
Qwen3.6 35B-A3B35B (3B active)Q3_K_M19 GB~58 tok/sRuns well
GLM-4.5-Air106B (12B active)Q4_K_M68 GB~7.4 tok/sPartial offload
Llama 4 Scout109B (17B active)Q4_K_M70 GB~5.3 tok/sPartial offload
gpt-oss-120b117B (5.1B active)MXFP468 GB~18 tok/sPartial offload
Qwen3.5 122B-A10B122B (10B active)———Won't fitWhat it needs
DeepSeek V4 Flash284B (13B active)———Won't fitWhat it needs
GLM-5.3-Flash320B (18B active)———Won't fitWhat it needs
GLM-5.3744B (40B active)———Won't fitWhat it needs

Performance

Tesla P40 vs. alternatives on Qwen3.8 27B

Generation speed in tokens per second for a single user. Token generation is limited mostly by memory bandwidth, which is why bandwidth predicts these results so well.

  1. RTX 3090

    Q4_K_M · 24GB

    ~39tok/s

  2. V100 32GB

    Q6_K · 32GB

    ~25tok/s

  3. MI50 32GB

    Q6_K · 32GB

    ~19tok/s

  4. Tesla P40

    Q4_K_M · 24GB

    ~11tok/s

Estimated (method)

Measured

Tesla P40 benchmark results

ModelQuantRuntimeGenerationPromptSource
Llama 2 7B ChatQ4_0llama.cpp (Vulkan)59 tok/s490 tok/sllama.cpp discussion #10879: Performance of llama.cpp with Vulkan · October 2026
Llama 2 7B ChatQ4_0llama.cpp (CUDA)55 tok/s1005 tok/sllama.cpp discussion #15013: Performance of llama.cpp on NVIDIA CUDA · October 2026

Our take

Should you buy the Tesla P40 for AI?

The Tesla P40 is a 2016 Pascal datacenter card with 24GB of GDDR5 on a 384-bit bus for 346 GB/s. It has no tensor cores and very slow FP16, so llama.cpp runs it through FP32 code paths and prompt processing is far slower than on modern cards. The card is passively cooled for server airflow, so desktop builds need a blower or 3D-printed fan shroud, and it takes an 8-pin EPS power connector rather than a PCIe plug. Used cards sold for about $230 to $330 in October 2026. CUDA 13 no longer supports Pascal, though the R580 driver branch keeps it working for now.

What we like

  • 24GB for the lowest price of any card
  • Works with llama.cpp and other CUDA 12 builds
  • Easy to buy in quantity

Watch out for

  • Slow FP16 and no tensor cores
  • Needs added cooling and an EPS power adapter
  • Dropped from CUDA 13

Best for: Very low-budget 24GB or 48GB builds · Home servers with strong airflow

Specifications

Tesla P40 specs that matter for AI

Memory
24GB GDDR5
Memory bandwidth
346 GB/s
Memory bus
384-bit
Architecture
Pascal
Power
250W
Interface
PCIe 3.0 x16
Released
2016
Street price (used)
~$230 (2026-10)
Software
CUDA, Vulkan

Sources: NVIDIA Tesla P40 product brief (PDF), Flopper.io Tesla P40 spec sheet

Cross-shop

Alternatives to the Tesla P40

NVIDIA2020
24GB

GDDR6X · 936 GB/s

RTX 3090

~$1,050 used

AMD2018
32GB

HBM2 · 1,024 GB/s

MI50 32GB

~$600 used

NVIDIA2018
32GB

HBM2 · 900 GB/s

V100 32GB

~$700 used

Frequently asked questions

How much VRAM does the Tesla P40 have?

The NVIDIA Tesla P40 has 24GB of GDDR5 with 346 GB/s of memory bandwidth on a 384-bit bus.

What is the largest LLM the Tesla P40 can run?

Fully in GPU memory, the largest model in our catalog that runs comfortably is Qwen3.6 35B-A3B (35B (3B active)) at Q3_K_M, using about 19GB with 8K context at ~58 tokens/sec.

Can the Tesla P40 run Llama 3.3 70B?

Only with partial CPU offload. At Q4_K_M the model needs about 47GB with 8K context, so part of it spills into system RAM and speed drops to about 1.6 tokens/sec with CPU offload. For usable 70B speeds you want 48GB of VRAM or more (for example two 24GB cards).

What is the best LLM to run on the Tesla P40?

Our top pick is Qwen3.8 27B at Q4_K_M, which fits in 24GB with room for 8K of context and runs at ~11 tokens/sec. See our full list of the best models for the Tesla P40 for coding, reasoning, and vision picks.

Is the Tesla P40 good for local AI in 2026?

The cheapest 24GB card, if you can cool it and accept Pascal's slow compute.