TL;DR
AMD Helios includes 72 MI455X GPUs, 31TB HBM4, and 2.9 exaflops of inference in a single rack. Built on open standards. Engineering samples second half of 2026, mass production second quarter of 2027.
AMD 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 response to Nvidia’s Vera Rubin NVL72. The system uses 18 compute trays, each with four MI455X accelerators on the new CDNA 5 architecture and a sixth-generation EPYC.Venice“CPU. Engineering samples are shipped in the second half of 2026. Mass production begins in the second quarter of 2027.
Architecture’s bet is on open standards. Helios uses UALink for expanded in-rack GPU interconnection, Ultra Ethernet Consortium specifications for scalable inter-rack networking, and the OCP Open Rack Wide form factor. Nvidia’s competing NVL72 uses proprietary NVLink. AMD is betting that data center operators who don’t want to be locked into a single provider’s interconnection will pay for the flexibility. AMD Pensando AI NICs drive the network with programmable hardware and UEC-ready RDMA.
Numbers are designed to compete in memory, not just computation. Each MI455X GPU carries HBM4 with 19.6 TB/s of bandwidth. The full rack offers 260 TB/s of scalable bandwidth and 43 TB/s of scalable bandwidth. That memory capacity is important for frontier model training and long-context inference, where the bottleneck has shifted from raw computing to the amount of data the system can hold and move. AI-driven memory crisis has sharply raised HBM pricesand 31 TB of HBM4 in a single rack represents a huge materials cost that only sovereign computing and hyperscaler budgets can absorb.
The 💜 of EU technology
The latest rumors from the EU tech scene, a story from our wise founder Boris and some questionable AI art. It’s free, every week, in your inbox. Register now!
Supermicro showed off the Helios hardware at Computex in June. AMD committed billions to UK AI infrastructure at London Tech Weekand Helios is the hardware on which those commits will run. The ROCm software stack supports PyTorch, TensorFlow, and JAX, meaning developers don’t need to rewrite code to leave Nvidia’s CUDA ecosystem, at least in theory. Whether AMD can close the software gap that has kept it behind Nvidia in AI computing is the question Helios is designed to force. Hardware specifications are competitive. The ecosystem is the proof.
