AMD has unveiled its latest ambitious AI hardware milestone with the Helios, a rack-scale system designed to directly challenge Nvidia’s dominant NVL72. The Helios system packs an impressive 72 Instinct MI455X GPUs, 31 terabytes of HBM4 memory, and delivers 2.9 exaflops of FP4 inference compute power — all housed within a single rack.
Building on CDNA 5 Architecture
At the heart of Helios lies the new CDNA 5 architecture, which powers the 18 compute trays that make up the system. Each tray contains four MI455X accelerators, offering a significant boost in performance and efficiency compared to previous generations. This marks AMD’s first foray into rack-scale AI computing, a domain that has largely been dominated by Nvidia’s offerings.
Strategic Move in the AI Race
The Helios system is a clear signal of AMD’s intent to compete head-on with Nvidia in the high-performance AI computing space. While Nvidia’s NVL72 has set the bar for large-scale AI infrastructure, AMD’s new system aims to offer a compelling alternative with its own strengths in memory bandwidth and compute density. Industry analysts are watching closely to see how Helios performs in real-world applications and whether it can carve out a significant market share in the enterprise AI segment.
The system’s launch underscores a broader shift in the tech industry toward modular, scalable AI infrastructure, where companies are increasingly looking for solutions that offer both performance and flexibility. With Helios, AMD is not just competing on raw compute power — it’s positioning itself as a serious contender in the next wave of AI hardware evolution.
