finds.dev

soumith/convnet-benchmarks

Easy benchmarking of all publicly accessible implementations of convnets

View on GitHub

★ 2,688PythonMITupdated Jun 2017

// the take

A benchmark suite comparing forward/backward pass speed of convolution kernels across Caffe, Torch, TensorFlow, Theano, Chainer, cuDNN, and a half-dozen other frameworks circa 2014-2015. Useful for anyone doing historical research into GPU convolution performance or curious how the deep learning framework landscape looked before cuDNN standardized things.

Actually reproducible: each framework directory ships the exact script, prototxt, or config used to generate its numbers, plus raw output logs, not just a table someone typed up. The layer-wise breakdown (L1-L5) isolates where time goes for kernels of different sizes and channel counts, which is more informative than a single end-to-end number. Comparing native Caffe/Torch conv against fbfft and cuDNN R2/R4 side by side captures a real inflection point in how convolutions were computed.

Last commit is from mid-2017 and the hardware is a Titan X with CUDA-era cuDNN R2/R4 — none of this reflects Ampere/Hopper, Tensor Cores, or any cuDNN version from the last eight years, so the numbers are historical curiosities, not a buying guide. Half the frameworks benchmarked (Chainer, Nervana neon, cuda-convnet2, DeepCL, Caffe-CLGreenTea) are dead projects, so most of the comparison table is now benchmarking vaporware against vaporware. No CI, no way to re-run this against a current GPU or current framework versions — it's a frozen snapshot with no path to becoming a living benchmark again.

We read the README, the directory tree and a key source file in September 2026. Nobody pays to be listed. How we judge →

// at a glance

Stars
2,688
Forks
571
Language
Python
License
MIT
Last commit
9 Jun 2017

// nearby

If convnet-benchmarks isn't quite it.

All repos →