AI Infrastructure 

AMD Helios is a rack-scale compute platform purpose-built for next-generation AI factories. The system integrates 72 AMD Instinct™ MI455X GPUs, AMD EPYC™ Venice processors, and AMD Pensando™ Vulcano AI NICs to eliminate bottlenecks and maximize data movement and system efficiency.
With industry-leading memory capacity and bandwidth, Helios seamlessly handles larger models and extended context windows while delivering ultra-responsive performance for complex multi-agent workflows. Built on open standards including OCP, ORW, UALink™, Ethernet/UEC, Helios enables flexible scaling across data centers while optimizing power delivery, liquid cooling, and serviceability—serving as the ideal foundation for scaling industrial-grade AI.

• Leadership Compute Performance

  • Integrates 72 AMD Instinct™ MI455X GPUs into a unified in-rack compute domain, delivering the scale and flexibility needed for inference, training, fine-tuning, and multi-agent workloads.

  • Delivers up to 2.9 EFLOPS of OCP MXFP4 and up to 1.4 EFLOPS of OCP MXFP6/MXFP8 performance.

• Industry's Best Memory

  • Large HBM memory capacity, implemented as a single load-store domain in the rack, helps support high-concurrency needs of agentic AI workloads.

  • Up to 31 TB HBM4 per rack with up to 1.67 PB/s aggregate memory bandwidth per rack.

• Open Scale-Up and Scale-Out Architecture

  • Scale-Up: UALink over Ethernet (UALoE) connects GPUs through the in-rack switch fabric (single-switch-hop, multi-plane fabric, blind-mate rear cable cartridge; 6 AMD Helios switch trays).

  • Scale-Out: Standards-based Ethernet scales across racks and clusters; Ethernet/UEC-aligned scale-out via AMD Pensando "Vulcano" AI NICs; up to 3x800 Gb/s per GPU, or up to 43 TB/s rack bidirectional bandwidth.