What it does

This repo benchmarks nine classic CNN architectures—AlexNet through ResNet-200—on CPUs and four NVIDIA GPUs, with and without cuDNN. All tests use a fixed batch size of 16 and 224×224 images, so the tables are directly comparable. The numbers are forward-plus-backward pass times in milliseconds.

The interesting bit

The README doubles as a hardware buying guide from 2016. The author draws explicit conclusions: Pascal Titan X beats GTX 1080 by ~1.4×, GTX 1080 edges out Maxwell Titan X by ~1.1×, and cuDNN provides 2–3× speedups across the board. A Pascal Titan X with cuDNN is stated to be 49–74× faster than dual Xeon E5-2630 v3 CPUs. These claims are grounded in the tables, not hand-waving.

Key highlights

• Covers AlexNet, Inception-V1, VGG-16/19, and ResNet-18/34/50/101/152/200
• Tests Pascal Titan X, GTX 1080, GTX 1080 Ti, Maxwell Titan X, and dual Xeon CPUs
• Separate forward and backward pass timings for each GPU/cuDNN combination
• Includes Top-1/Top-5 error rates for accuracy-speed tradeoff comparisons
• Model files and conversion scripts provided (2.1 GB download)

Caveats

• All benchmarks run in Torch, not modern PyTorch or TensorFlow
• CUDA 8.0 and cuDNN 5.0/5.1 are ancient; current speedups likely differ
• ResNet-200 failed on the 8 GB GTX 1080 due to memory limits
• VGG-16 and VGG-19 use dense prediction (256×256), giving them a slight accuracy advantage versus single-crop for other models

Verdict

Worth a look if you’re researching historical hardware scaling laws or defending a 2016 GPU purchase. Skip it if you need current PyTorch benchmarks—this is a period piece, not a living benchmark suite.