GPU Collective Latency Optimized to Near Speed-of-Light, Boosting LLM Inference

traviscline · x · 2026-08-25

This arXiv paper investigates optimizing GPU collective communication latency to approach the hardware Speed-of-Light (SoL) limit within scale-up networks. Emerging workloads like long-context LLM inference are increasingly latency-bound. The authors developed low-latency interfaces based on NCCL's device-side API, implementing barrier-free synchronization and efficient symmetric memory usage. Microbenchmarks show substantial latency reductions for small messages, with overhead within 7% of the SoL limit. Integration into real applications improved inter-token latency and throughput for LLM inference and accelerated cuSOLVERMp.

Original post →

More from Infra

Infra channel →