1. Learn
  2. Pick your hardware

The Tesla P100 for local LLMs: 16 GB of HBM2 for under $100

What we learned running a rack of 2016 data-center P100s for AI. Real speeds, what it takes to cool and power them, and the software catch.

Updated Oct 26, 2026 · RunItLocal.io

The Tesla P100 is a data-center card from 2016. It has no fans, no display outputs and no tensor cores. It's also one of the best deals in local AI, because it has 16 GB of HBM2 memory at 732 GB/s, and used ones sell for under $100. We run a rack of them.

Why it's fast at this

Generating text is limited by memory bandwidth, not compute (see why bandwidth sets speed). The P100's bandwidth is more than double that of a new RTX 4060 Ti, so for generating text it's simply faster.

Tesla P100, 16 GB · estimated with our calculator · full page

  • Qwen3 4BQuick everyday chat and summaries
    ~160 tok/s
    Runs greatqwen3:4b
  • Qwen3 8BA solid all-rounder for chat and writing
    ~77 tok/s
    Runs greatqwen3:8b
  • Gemma 3 12BReads images as well as text
    ~50 tok/s
    Runs greatgemma3:12b
  • Qwen3 14BNoticeably smarter at reasoning and code
    ~43 tok/s
    Runs greatqwen3:14b
  • gpt-oss 20BOpenAI's open model. Fast, strong reasoning
    ~130 tok/s
    Runs greatgpt-oss:20b
  • Gemma 3 27BGoogle's best open model for its size
    needs 19 GB
    Too biggemma3:27b
  • Qwen3 30B-A3B30B smarts at small-model speed
    needs 21 GB
    Too bigqwen3:30b
  • Qwen3 32BStrong coding and long answers
    needs 22 GB
    Too bigqwen3:32b
  • Llama 3.3 70BClose to the big cloud models
    needs 45 GB
    Too bigllama3.3:70b
  • gpt-oss 120BOpenAI's largest open model
    needs 67 GB
    Too biggpt-oss:120b
  • Qwen3 235B-A22BAlibaba's flagship open model, for multi-GPU rigs
    needs 144 GB
    Too bigqwen3:235b
  • DeepSeek R1 671BThe full DeepSeek R1, for the biggest rigs
    needs 406 GB
    Too bigdeepseek-r1:671b

What we measured

On our rack, two P100s run Qwen3.6-35B-A3B at 53 tok/s. That's a 35B mixture-of-experts model split across both cards' 32 GB. In a run of a billion tokens across 90 used GPUs, the P100s never throttled, and power came to about 6 to 8 cents per million tokens.

The catches

Cooling is not optional. The P100 is passive: it expects a server pushing air straight through it. In a normal case it overheats within minutes. You need either a server chassis or a 3D-printed or bought blower shroud on each card, about $15 to $25 each.

Power uses a CPU-style plug. The P100 takes an 8-pin EPS connector, like a motherboard's CPU power, not a normal graphics card plug. Use the right adapter. Plugging in the wrong cable can damage the card or the supply.

Old architecture, limited software. Pascal (compute capability 6.0) is supported by llama.cpp and the tools built on it, including Ollama and LM Studio. Many newer server engines, like vLLM, need newer cards. NVIDIA's newest CUDA releases have also dropped Pascal, so stay on a driver and CUDA version that still supports it.

Slow at reading long prompts. No tensor cores means prompt processing is much slower than on an RTX card. Chatting is fine; pasting a 50-page document takes a while.

No display output. You need another GPU, or the CPU's graphics, for a monitor.

What to build around it

A used workstation like a Dell T7810 or HP Z840 is ideal: lots of PCIe slots, big power supplies, and 64 GB of RAM for a few hundred dollars. Two P100s plus shrouds and cables come to about $450 to $650 all in. That's our "Best value" build on the builds page, with a full parts list.

Who it's for

  • Good fit: tinkerers who want 32 GB of fast memory cheaply, homelabs, and anyone serving a model to a few people around the clock.
  • Bad fit: a quiet desk machine, image generation (slow without tensor cores), or anyone who wants it to just work out of the box. For that, a used RTX 3090 is the easy path.

Running one already? Put your numbers on the leaderboard.

Models in this guide