SURMA: Efficient Fine-Grained Communication Offload for SuperPOD
Haoran Shang, Weifeng Zhang, Mengzhu Li, Huizheng Wang, Sihan Guan, Xinyang Huang, Bowen Liu, Ran Shu 等 17 位作者
原文摘要
GPU SuperPODs underpin large language model (LLM) serving, but growing bandwidth and round-trip times make inter-GPU communication consume increasing compute resources. SM-driven remote memory access occupies shared memory or registers for in-flight data and competes for SM-to-NoC bandwidth. Dedicated engines can free these resources, but must handle fine-grained transfers, down to a 576\,B KV-cache entry. NVIDIA's CopyEngine uses less than 1\% of NVLink bandwidth at 7\,KB granularity. We present SURMA, an on-die Scale-Up Remote Memory Access engine. Adapting RDMA requires overcoming slow doorbell submission, excessive per-application connection state, head-of-line blocking of completion notifications on shared connections, and costly ordering for multipath transfers. SURMA uses a credit-based shared send queue to reduce submission latency from 1.2\,$μ$s to 400\,ns; fused write-with-atomic semantics to share connections while independently notifying each transaction; and Network Per Path Ordering to ensure data arrives before the completion flag without Sender or Receiver Ordering. We implement SURMA in RTL on a two-FPGA prototype and evaluate it with two 8-GPU NVLink servers. At 512\,B granularity, SURMA saturates attainable bandwidth with a single SM, whereas NVLink needs over 14\% of the SMs. SURMA occupies under 0.5\% of GPU area and accelerates inference on four models by 1.05--1.26$\times$ over NVLink.
cs.AR