Lightning Talk: Standardizing CPU Benchmarking with TorchBench for PyTorch...
summary
TorchBench now includes a standardized CPU harness that runs FX INT8 and torch AMP (bf16/fp32) on CPU, compares channel_last (NHWC) vs channel_first formats, and adds GNN coverage (GCN, GraphSAGE) via curated GNN subsets to avoid huge datasets. The repo implements a userbenchmark CPU harness with core binding, NUMA and thread-affinity controls, measures throughput and latency, and is already able to detect regressions and demonstrate concrete speedups for CPU-specific optimizations.
AI-written summary · about
related talks
17mLightning Thunder: Supercharged PyTorch for Modern Hardware
08mLightning Talk: Achieving SOTA GEMM Performance: A CuTeDSL Backend for PyTorch Induc...
22mBridging the Hardware Gap With Code Harnesses on the Hugging Face Kernels Hub
11mSotabench for Reproducible Research
10mPyTorch Performance
13mLightning Talk: Challenges and Standardization in PyTorch Ecosystem...
24mAdvancements in High-Performance AI/ML through PyTorch's Python Compiler
14mLightning Talk: Accelerating PyTorch Models With Torch.compile's C++ Wrapper Mode