TL;DR
AMD Helios packs 72 MI455X GPUs, 31TB HBM4, and 2.9 exaflops of inference into one rack. Built on open standards. Engineering samples H2 2026, mass production Q2 2027.
AMD’s Helios is a single rack containing 72 Instinct MI455X GPUs, 31 terabytes of HBM4 memory, and 2.9 exaflops of FP4 inference compute. It is AMD’s first rack-scale AI system and its direct answer to Nvidia’s Vera Rubin NVL72. The system uses 18 compute trays, each holding four MI455X accelerators on the new CDNA 5 architecture and one sixth-generation EPYC “Venice” CPU. Engineering samples ship in the second half of 2026. Mass production begins Q2 2027.
The architecture bet is open standards. Helios uses UALink for scale-up interconnect between GPUs within the rack, Ultra Ethernet Consortium specifications for scale-out networking between racks, and the OCP Open Rack Wide form factor. Nvidia’s competing NVL72 uses proprietary NVLink. AMD is betting that data centre operators who do not want to be locked into a single vendor’s interconnect will pay for the flexibility. AMD Pensando AI NICs handle the networking with programmable hardware and UEC-ready RDMA.
The numbers are designed to compete on memory, not just compute. Each MI455X GPU carries HBM4 with 19.6 TB/s of bandwidth. The full rack delivers 260 TB/s of scale-up bandwidth and 43 TB/s of scale-out bandwidth. That memory capacity matters for frontier model training and long-context inference, where the bottleneck has shifted from raw compute to how much data the system can hold and move. The AI-driven memory crisis has pushed HBM prices up sharply, and 31TB of HBM4 in a single rack represents an enormous materials cost that only hyperscaler and sovereign compute budgets can absorb.
The 💜 of EU tech
The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!
Supermicro showed Helios hardware at Computex in June. AMD committed billions to UK AI infrastructure at London Tech Week, and Helios is the hardware those commitments will run on. The ROCm software stack supports PyTorch, TensorFlow, and JAX, which means developers do not need to rewrite code to move from Nvidia’s CUDA ecosystem, at least in theory. Whether AMD can close the software gap that has kept it behind Nvidia in AI compute is the question Helios is designed to force. The hardware specs are competitive. The ecosystem is the test.
