Leaderboard
Measured on real hardware by the people who ran it, from each device's own energy meter. Nothing here is a sample or an estimate, and nobody's name is shown: a person only counts towards the number of people behind a row.
GPUs: work per joule
Every row is the same test: 8192 × 8192 matrix multiplications, 20 seconds in fp16 and 20 in fp32 after a warm-up, with the energy from the GPU's own meter. GFLOP per joule is the work the card does for each joule it draws; higher is better. The power limit and throttling are shown because they decide much of the result: a T4 in a notebook usually runs against its 70 W limit the whole time.
Loading…
Put your GPU on the board
- Create a free account, then an API key under Settings → API keys.
- In a Kaggle or Colab notebook with a GPU switched on, or in a terminal in the Python environment that has PyTorch (any NVIDIA or AMD GPU):
pip install -U landauer-gapjoules login lgk_…In a notebook, start each line withjoules bench gpu --share!. The test takes about a minute and prints the link to its row.
joules sends only the GPU's name, its power limit, and per precision the number of multiplications, the energy, the time and how much the card throttled. Nothing from your notebook, files or machine name.
Local models: energy per token
Machines are ranked against each other on the same model, never one model against another: a smaller model always uses less energy per token, which does not make it the better choice. Each row is energy above idle, and says what its meters cover; a Mac's meter is the whole chip, an NVIDIA one the GPU alone. Each person's runs count once.
Loading…
Put your machine on the board
- Create a free account, then an API key under Settings → API keys.
- Install joules and Ollama (ollama.com/download), then benchmark a model:
curl -LsSf https://wattops.dev/install.sh | sh && export PATH="$HOME/.local/bin:$PATH"On Windows, install withjoules bench ollama:llama3.2:3b --shareirm https://wattops.dev/install.ps1 | iexin PowerShell. joules asks for your key once, measures the idle machine for 5 seconds, and downloads the model the first time.
joules sends only the model name, hardware, energy per token, speed and average power. Nothing from your prompts, outputs, host name or files.
A free account adds your machines to these boards, keeps every run you measure on your own Tracking page, and lists your results so you can remove them any time.
Create a free account