Performance Evaluation of Interconnect Technologies for AI Scale Up Computing UAL vs UALoE SU скачать в хорошем качестве

Performance Evaluation of Interconnect Technologies for AI Scale Up Computing UAL vs UALoE SU

Srihari Vegesna - , Srinivas Gangam FELLOW - Upscale AI AI workloads increasingly demand low-latency, high-bandwidth interconnects. This paper compares Ultra Accelerator Link (UAL), Ethernet-based UALoE-SUE, and RoCE for AI-optimized data transfer. We evaluate transaction-level load balancing, DMA thread efficiency, and bandwidth utilization for memory-semantic XPU workloads. UAL outperforms in small transfer scenarios, leveraging deterministic latency and memory transaction-based load balancing. Ethernet protocols, in contrast, suffer from variable latency and inefficient handling of small packets, limiting their effectiveness in scale-up AI systems.