AMD’s Helios rack uses 72 MI455X GPUs and 260 TB/s of scale-up bandwidth
SumitGup · x · 2026-07-25
AMD's Helios rack-scale platform ties 72 MI455X GPUs into one shared-memory system
The post breaks down AMD’s Helios rack-scale AI architecture, which uses 72 MI455X GPUs, 12 Broadcom Tomahawk 6 ASICs, and 36 parallel network planes to deliver 3.6 TB/s bidirectional bandwidth per GPU and 260 TB/s aggregate scale-up bandwidth.
- UALoE defines GPU-to-GPU communication semantics, including memory reads and writes, ordering, and traffic distribution across the 36 planes.
- ESUN provides the Ethernet fabric underneath, handling forwarding, flow control, congestion management, retries, and small-message transport.
- The result is a rack where 31 TB of HBM4 remains physically distributed but can be accessed as if the whole cabinet were a single tightly coupled computer.
The image also shows the scale-up architecture as an Ethernet-based stack built around UALoE, OCP SUE, and ESUN, with AMD, Broadcom, Cisco, Meta, Microsoft, NVIDIA, and others listed as founding members.
Related event: AMD Launches MI455X and Helios, Escalating AI Compute to Rack-Scale(9 posts)→
More from Infra
- WSJ: Nvidia is in talks to backstop about $250 billion of OpenAI's data center plan — KateClarkTweets · 2026-07-27
- YC talk on BCI x AI says infrastructure is what really determines speed — garrytan · 2026-07-27
- A 13B model ran on a no-GPU PC by paging weights from SSD via llama.cpp — ID_R_McGregor · 2026-07-27
- llama.cpp warns that GGUFs made before a recent change must be regenerated — EconomySerious · 2026-07-27
- RTX 5090 local tests show Qwen Q6 can drop to 15 tok/s at 80k context — LFAdvice7984 · 2026-07-27
- Surprising Ubuntu Setup: NVIDIA 5090 PC Becomes the Easiest AI Rig — _xjdr · 2026-07-27