AMD's Helios AI Rack Unveiled: MI455X GPUs Meet Venice CPUs in a 72-GPU, Open-Standards AI Powerhouse
AMD revealed the Helios AI Rack, a full-stack, liquid-cooled AI platform pairing 72 Instinct MI455X GPUs with 6th Gen AMD EPYC Venice CPUs on a 2nm process, plus Pensando networking (Vulcano 800 NIC and Salina DPU) and ROCm software. It promises massive FP4/FP8 compute (up to about 2.9 EFLOPS FP4 and 1.4 EFLOPS FP8), 31 TB of HBM4, and high-bandwidth interconnects designed to rival NVIDIA’s rack-scale solutions, with Azure-based deployments and early customers like Microsoft, OpenAI, Meta, and Oracle. Price is around $5–5.5 million per rack and power draw around 225–245 kW.