GPU sidecar routed my ASIC proxy in under 10 seconds
The OpenROAD GPU Sidecar is an experimental tool designed for placement, global routing, and detailed routing in ASIC design. Performance tests on the HC-1 proxy design showed rapid execution times for each stage, with total GPU work measured in milliseconds. This sidecar operates independently of the OpenROAD project, utilizing a unique interchange format for design data.
- ▪The GPU Sidecar achieved placement in 1.00 second, global routing in 0.93 seconds, and detailed routing in 5.10 seconds.
- ▪It is designed to work with OpenROAD-compatible design data, exporting and importing using the ORAX format.
- ▪The tool is not a patch to OpenROAD but an experimental accelerator that runs outside the OpenROAD source tree.
Opening excerpt (first ~120 words) tap to expand
OpenROAD GPU Sidecar CUDA sidecar for OpenROAD-style placement, global routing, and detailed routing. This is an experimental Acculux Systems sidecar. It is not affiliated with, endorsed by, or merged into the OpenROAD project. Performance Measured on the HC-1 proxy design on a RunPod NVIDIA GeForce RTX 4070 Ti: Stage Runtime Placement 1.00s Global routing 0.93s Detailed routing 5.10s Profiled detailed-route run on the same pod: Detailed-route wall time: 3.80s Total CUDA-timed GPU work: 10.377 ms Placement CUDA-timed GPU work: 30.252 ms Global-route CUDA-timed GPU work: 16.508 ms This is the working sidecar path validated on the pod. The remaining wall-clock overhead is mostly host-side I/O and allocation work, not the routing kernels themselves.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at GitHub.