CCL-Bench — Benchmark Results

CCL-Bench is a collaborative benchmarking project studying how distributed ML workloads behave across different hardware, frameworks, and communication libraries. As models scale, a distributed implementation on hardware A with library X can behave drastically differently from hardware B with library Y. Each group collects traces and contributes analysis tools, which are then applied cross-group — making the benchmark scalable without brute-force exploration. Metrics include MFU, estimated memory bandwidth, step time, and more.

Part of the Cornell sysphotonics research group  ·  Maintainers: Eric Ding, Kaiwen Guo, Jelena Gvero, Byungsoo Oh, Bhaskar Kataria, Atharv Sonwane, Rachee Singh  ·  Contact: Eric Ding  ·  Contributing: GitHub