AMD reveals Helios rackscale system architecture, unifying server hardware and AI data center infrastructure into a cohesive design.

AMD's announcement of its Helios rack at this year's Advancing AI event represents the company's most significant product reveal to date. Helios is a rackscale system—a computing appliance built by integrating AMD EPYC CPUs, Instinct MI455X GPUs, and Pensando DPUs/NICs alongside ZT Systems engineering to deliver performance that exceeds what individual servers can achieve alone.

Rackscale systems have become essential infrastructure for modern datacenters as AI workloads grow beyond the capabilities of single-server architectures. These systems cluster multiple heterogeneous servers into a unified computing platform, addressing both the processing power and memory capacity demands of large-scale AI. This paradigm has driven NVIDIA's dominance in GPU computing, and it now represents AMD's opportunity to establish competitive parity.

The engineering challenge extends beyond simply combining components. A functional rackscale AI system requires high-performance fabrics to interconnect CPUs, GPUs, and other processors. The critical advances enabling rackscale systems have centered on integrating these fabrics into GPUs themselves and then designing networking hardware to extend those fabrics across multiple nodes within a single rack.

AMD's path to Helios accelerated after acquiring Pensando Systems in 2022. Four years later, all necessary pieces finally converged. The resulting system—combining AMD EPYC 9006 "Venice" CPUs, Instinct MI455X GPUs, and Pensando NICs/DPUs—represents a flagship technical and business achievement for the company.

The production Helios specifications closely match AMD's original projections. The 72-GPU configuration delivers 2.9 EFLOPS of peak MXFP4 compute performance. Each GPU pair connects to 432GB of HBM4 memory with 23.3TB/second of bandwidth, yielding a cumulative system capacity of 31TB of HBM4 memory and 1.7PB/second of total memory bandwidth—approximately 21% higher than originally projected. This bandwidth improvement addresses a critical bottleneck in AI inference.

On the networking side, Helios incorporates dual domains: scale-up networking for intra-rack connectivity provides 260TB/second of cumulative bandwidth between nodes, while scale-out networking enables multi-rack clustering at 43TB/second. These represent unprecedented bandwidth capabilities from a single rack system.