AMD’s Helios: A Rack-Scale AI System with 72 GPUs and 31TB of HBM4
AMD’s Helios packs an impressive 72 MI455X GPUs, 31 terabytes of HBM4 memory, and 2.9 exaflops of FP4 inference into a single rack. It is AMD’s first rack-scale AI system and a direct competitor to Nvidia’s Vera Rubin NVL72.
Key Features:
- Performance: Delivers 2.9 exaflops of FP4 inference and 1.4 exaflops of FP8 training.
- Open Standards: Built on open standards, including UALink for scale-up interconnect and Ultra Ethernet Consortium specifications for scale-out networking.
- Memory: Each MI455X GPU features HBM4 with 19.6 TB/s of bandwidth, providing a massive memory capacity for frontier model training and long-context inference.
- Ecosystem: Supports PyTorch, TensorFlow, and JAX, aiming to bridge the software gap with Nvidia’s CUDA ecosystem.
Launch and Availability:
- Engineering Samples: Expected to ship in H2 2026.
- Mass Production: Scheduled to begin Q2 2027.
The rackscale system is a significant move by AMD into the AI hardware market, targeting data center operators seeking flexibility over interconnect options. With its high memory capacity and open standards approach, Helios aims to challenge Nvidia’s dominance in AI compute.