AMD is targeting the AI accelerator market with rackscale integrated infrastructure aimed at the frontier AI era. AMD said it unveiled the 'AMD Helios' rackscale solution at its 'Advancing AI 2026' event in San Francisco on July 23 and has begun mass production.

AMD Helios is an AI infrastructure platform that integrates AMD Instinct MI455X GPUs, sixth-generation AMD EPYC server CPUs, AMD ROCm software and AMD Pensando networking into a single rackscale architecture. On a per-rack basis, it provides up to 2.9 exaflops peak FP4, 31 terabytes of HBM4 memory and 1.7 petabytes per second of memory bandwidth. FP4 is a 4-bit floating-point format used for AI computing, a method that boosts processing speed and energy efficiency in exchange for lower precision.

AMD said that, compared with rival flagship solutions, peak FP4 performance is 15 percent higher, HBM capacity is 50 percent higher, HBM bandwidth is 6 percent higher and scale-out bandwidth is 50 percent higher. The company said token processing costs per rack are also up to 30 percent lower than competing solutions.

In terms of scalability, AMD Helios can be configured from a single rack to a gigawatt-class AI cluster. A single rack includes 18 4-GPU compute trays, for a total of 72 GPUs, and the AMD Pensando Vulcano 800 AI NIC handles the high-bandwidth, low-latency connections needed for distributed inference and large-scale model training. It uses an open architecture, supporting UALink over Ethernet-based scale-up and Ultra Ethernet Consortium standard-based scale-out.

AMD Helios will be used by major AI companies such as Meta, OpenAI, Anthropic and Microsoft to build gigawatt-class AI infrastructure. "AMD Helios has brought together leadership compute, high-performance networking and open software into an integrated rackscale platform," said Vamsi Boppana (밤시 보파나), senior vice president of AMD's AI unit. "It gives customers the flexibility to accelerate large-scale inference, frontier model training and next-generation AI infrastructure buildouts," he said.