Tomahawk 6
Search documents
Ethernet Scale Up Networking for Next Generation AI Workloads.
AMD· 2026-08-07 13:59
System Architecture & Hardware Specifications - AMD Helios architecture integrates 72 MI455 GPUs per rack, supported by 18 compute nodes with each node containing 1 CPU and 4 GPUs[8][64] - Scale-up interconnect utilizes 12 Broadcom Tomahawk 6 switch ASICs (A6) connected to each GPU via three 400-gigabit Ethernet links, creating 36 parallel network planes[11][12] - Tomahawk 6 chip series includes models with 1024 100-gigabit SerDes achieving 102 terabits per second, and 200-gigabit SerDes variants built on 3-nanometer technology[71][72][73] - Each MI455 GPU provides 23 terabytes per second of high-bandwidth memory (HBM) bandwidth and 3.6% terabytes per second of scale-up bandwidth[54][55] Network Technology & Performance - Scale-up network operates on Ethernet technology using UAL-OE (UALink over Ethernet) protocol, delivering 1.8 terabytes per second of bandwidth per GPU with single-layer switching topology[11][15][25] - Reliability and resilience mechanisms allow the system to survive the failure of any individual link, individual switch, or entire switch tray by redirecting traffic across parallel network planes[20][21] - Advanced network features include compressed headers for optimizing small message delivery, link layer retry for error handling without transport layer loss, and a fully unified packet buffer to absorb traffic bursts[76][79][81]