AMD's Helios integrates 72 GPUs and 31 terabytes of HBM4 within a single rack. This is AMD's response to Nvidia's NVL72.

      **TL;DR**: AMD's Helios includes 72 MI455X GPUs, 31TB of HBM4, and delivers 2.9 exaflops of inference in a single rack, built on open standards. Engineering samples expected in the second half of 2026, with mass production starting in Q2 2027.

      AMD's Helios is a rack containing 72 Instinct MI455X GPUs, 31 terabytes of HBM4 memory, and 2.9 exaflops of FP4 inference compute power. This marks AMD's inaugural rack-scale AI system and serves as a direct competitor to Nvidia's Vera Rubin NVL72. The design comprises 18 compute trays, with each tray accommodating four MI455X accelerators based on the new CDNA 5 architecture, along with one sixth-generation EPYC “Venice” CPU. Engineering samples are set to ship in the latter half of 2026, while mass production is slated for Q2 2027.

      The architecture emphasizes open standards, utilizing UALink for the interconnect between GPUs within the rack, adhering to Ultra Ethernet Consortium specifications for networking between racks, and incorporating the OCP Open Rack Wide form factor. In contrast, Nvidia’s NVL72 relies on proprietary NVLink. AMD believes that data center managers seeking flexibility and not wanting to be tied to a single vendor's interconnect will invest in this option. Networking is managed by AMD Pensando AI NICs with programmable hardware and UEC-ready RDMA capabilities.

      The specifications aim to compete on memory in addition to computing power. Each MI455X GPU features HBM4 with 19.6 TB/s bandwidth, resulting in a total rack bandwidth of 260 TB/s for scale-up and 43 TB/s for scale-out. This memory capacity is crucial for training advanced models and long-context inference, where the limitation has transitioned from sheer computational power to the amount of data the system can store and transfer. The AI-driven memory shortage has caused significant increases in HBM prices, and 31TB of HBM4 in a single rack entails a substantial material cost that only large-scale cloud providers and government computing budgets can accommodate.

      Supermicro unveiled Helios hardware at Computex in June. AMD has invested billions into AI infrastructure in the UK, as showcased during London Tech Week, and Helios will form the foundation of these initiatives. The ROCm software stack provides support for frameworks like PyTorch, TensorFlow, and JAX, implying that developers should, theoretically, be able to transition from Nvidia’s CUDA ecosystem without needing to rewrite code. The key question is whether AMD can overcome the software disparity that has historically placed it behind Nvidia in AI compute capabilities. While the hardware specifications are competitive, the broader ecosystem remains the real challenge.