AMD has unveiled Helios, a rack-scale AI system packing 72 Instinct MI455X GPUs, 31TB of HBM4 memory, and 2.9 exaflops of FP4 inference compute into a single rack. Built on 18 compute trays each combining four MI455X accelerators (CDNA 5 architecture) with a sixth-gen EPYC Venice CPU, Helios directly challenges Nvidia's Vera Rubin NVL72. A key differentiator is AMD's bet on open standards — UALink for GPU interconnect, Ultra Ethernet Consortium specs for scale-out networking, and OCP Open Rack Wide form factor — versus Nvidia's proprietary NVLink. The full rack delivers 260 TB/s of scale-up bandwidth and 43 TB/s of scale-out bandwidth. Engineering samples are expected in H2 2026, with mass production in Q2 2027. The ROCm software stack supports PyTorch, TensorFlow, and JAX, though closing Nvidia's CUDA ecosystem advantage remains the critical challenge.
