AMD Unveils Helios AI System with 72 GPUs and 31TB HBM4, Challenging Nvidia
AMD has introduced its Helios, a new rack-scale AI system designed to compete directly with Nvidia's offerings. This single rack houses an impressive 72 AMD Instinct MI455X GPUs, providing a substantial 2.9 exaflops of FP4 inference compute power. The system is further augmented by 31 terabytes of HBM4 memory, indicating a significant investment in high-bandwidth memory for AI workloads. Helios is structured using 18 compute trays, with each tray containing four MI455X accelerators built on AMD's CDNA 5 architecture. This launch marks AMD's entry into the high-density, rack-scale AI system market, positioning Helios as a direct competitor to Nvidia's Vera Rubin NVL72 system.
AMD's introduction of the Helios system signifies a strategic move to address the escalating demand for high-performance, integrated AI computing infrastructure. By consolidating 72 GPUs and extensive HBM4 memory within a single rack, AMD is directly confronting the architectural advantages Nvidia has established in this market segment. This development highlights a critical industry trend: the shift from discrete component sales to comprehensive, optimized system solutions for large-scale AI deployments. The competitive pressure between AMD and Nvidia in this domain is likely to accelerate innovation in GPU architecture, memory technology, and system integration, potentially leading to more efficient and powerful AI hardware solutions in the coming years. Companies will need to carefully evaluate the total cost of ownership, performance benchmarks, and ecosystem support when choosing between these high-density systems.
AI-generated to prompt reflection — not editorial opinion, not advice, not a statement of fact. How this works.