Optimizing GNNs for the Wild: PyTorch-to-ONNX Acceleration of GarNet on CPUs and GPUs
May 25, 2026·
·
0 min read
Uzziel Perez
Abstract
We benchmark GarNet graph neural networks in PyTorch against ONNX Runtime across multi-core CPUs and NVIDIA A100 GPUs, achieving up to 5× speedup on CPU and ~2× on GPU while maintaining FP32 numerical parity at the 10⁻⁷ level. The work targets deployment in the LHCb GPU high-level trigger (HLT1/Allen).
Date
May 25, 2026 — May 29, 2026
Event
Location
Chulalongkorn University, Bangkok, Thailand
