At Computex 2026, AMD unveiled its first rack-scale AI system, Helios, powered by Instinct MI455X accelerators and EPYC...

At Computex 2026, AMD unveiled its first rack-scale AI system, Helios, powered by Instinct MI455X accelerators and EPYC Venice CPUs, aiming to rival Nvidia's NVL72 VR200. However, initial systems rely on UALink-over-Ethernet interconnect rather than dedicated UALink, due to unavailability of UALink switches. The article Tom's Hardware highlights that Ethernet introduces higher latency, protocol overhead, and less deterministic performance compared to a dedicated scale-up fabric, potentially hamstringing real-world training performance. While Helios matches Nvidia's scale-up bandwidth on paper (260 TB/s aggregated) and offers 31 TB HBM4 memory capacity, the interconnect choice may limit its competitiveness in large-scale AI training. The systems are expected to be available later this year, but the reliance on a general-purpose networking technology designed for non-AI workloads raises concerns about AMD's ability to deliver competitive performance in the hyperscale AI market.