Blog
ResNet-50 max training batch: OpenNN vs PyTorch vs TensorFlow
This benchmark asks a capacity question, not a speed question: what is the largest fp32 training batch that completes one real ResNet-50/CIFAR-10 training step on the same GPU?…
OpenNN
Blog
Benchmarks, platform notes, and technical articles from the OpenNN team.
Blog
Capabilities: OpenNN vs PyTorch
OpenNN and PyTorch both build, train and run dense, convolutional, LSTM and Transformer networks on CPUs and NVIDIA GPUs. PyTorch offers more models and distributed training.
Blog
Inference performance: OpenNN vs PyTorch
Inference speed, peak memory and energy of dense, LSTM, CNN and Transformer networks on CPU and GPU. OpenNN led PyTorch in all 18 comparisons.
Blog
Startup time: OpenNN vs PyTorch
OpenNN applications delivered their first prediction 63% to 99.8% sooner after launch than the same applications written with PyTorch.
Blog
Deployment size: OpenNN vs PyTorch
An OpenNN application needs 65% to 99% less disk space than the same application on a standard PyTorch Python installation, without counting model weights.
Blog
Code size: OpenNN vs PyTorch
The OpenNN library has 96% fewer source lines than PyTorch, and a small classification application takes 9 lines with OpenNN versus 29 with PyTorch.
Blog
Prediction quality: OpenNN vs PyTorch
Trained from the same initial weights, OpenNN and PyTorch reached similar held-out scores with dense, LSTM, CNN and Transformer networks.
Blog
Training performance: OpenNN vs PyTorch
Training speed, peak memory and energy of dense, LSTM, CNN and Transformer networks on CPU and GPU. OpenNN led PyTorch in all 18 comparisons.