AMD Launches Helios, Exascale AI Infrastructure

AMD has entered the AI accelerator market with its rack-scale integrated infrastructure targeting the frontier AI era. At the 'Advancing AI 2026' event in San Francisco, AMD announced the launch of its Helios rack-scale solution, which combines AMD Instinct MI455X GPUs, 6th-gen AMD EPYC server CPUs, AMD ROCm software, and AMD Pensando networking into a single AI infrastructure platform. The Helios system provides up to 2.9 exaflops peak FP4 performance per rack, 31 terabytes of HBM4 memory, and 1.7 petabytes per second memory bandwidth. FP4 is a 4-bit floating-point format prioritizing speed and energy efficiency over precision. AMD claims Helios outperforms competing solutions by 15% in peak FP4 performance, 50% in HBM capacity, 6% in HBM bandwidth, and 50% in scale-out bandwidth. It also reduces token processing costs by up to 30% compared to rival solutions. Helios supports scaling from single racks to gigawatt-scale AI clusters, with 18 GPU compute trays (72 GPUs) per rack and high-bandwidth, low-latency connections via AMD Pensando Buracano 800 AI NICs. The open architecture supports UALink-over-Ethernet scaling and Ultra Ethernet consortium standards. Helios is set to be used by major AI firms including Meta, OpenAI, Anthropic, and Microsoft.