AMD Helios Platform and MI400 Series Accelerators Move AI Data Centers to Double-Wide Racks
At Computex 2026, AMD unveiled its next-generation AI compute roadmap, highlighted by the AMD Helios rack-scale AI platform and its upcoming Instinct MI400 series accelerators (including the MI430X, MI440X, and MI455X). Slated for deployment starting in late 2026, the Helios platform introduces a major physical design shift: the transition to a double-wide rack architecture.
This double-wide rack configuration is designed to handle the massive power density and scale of next-generation hardware. Forrest Norrod, General Manager of AMD's Data Center Solutions Business Group, confirmed that the Helios platform is a direct output of ZT Systems, the rack-scale engineering firm AMD acquired in late 2024.
Technical Specifications and Open Standards
The AMD Helios platform is built to deliver up to 10x more AI compute performance compared to the previous-generation MI355X. Key technical highlights of the Helios and MI400 architecture include:
- HBM4 Memory Integration: The platform will scale up to 31TB of HBM4 memory and 1.4PB/s of memory bandwidth per rack. This points to approximately 440GB of memory per GPU across 72 GPUs in a rack, utilizing 12 stacks of 36GB HBM4 memory (which Micron began shipping in early 2026).
- Next-Gen Processors: The rack integrates AMD's upcoming 2nm "Venice" EPYC CPUs, which feature up to 256 cores, PCIe Gen6 support, and 1.6TB/s of memory bandwidth.
- Networking and Interconnects: Helios will adopt open industry standards, utilizing the AMD Vulcano 800G NIC and incorporating UALink (Ultra Accelerator Link) and Ultra Ethernet Consortium (UEC) standards.
During the announcement, OpenAI CEO Sam Altman appeared on stage to highlight OpenAI's close collaboration with AMD on the Helios platform, demonstrating growing hyperscaler support for AMD as an alternative to NVIDIA.