AMD's Helios AI system is ready for exascale computing
AMD just dropped Helios, its new rack-scale AI system, at OCP Global Summit.
Built to handle the biggest AI projects out there, Helios delivers 1.4 exaFLOPS (FP8) of power thanks to 72 Instinct MI450 GPUs—all kept cool with liquid cooling tech.
It runs on AMD's open ROCm software stack
Helios packs 31TB of ultra-fast HBM4 memory and hits a huge bandwidth of 1.4PBps.
It uses advanced network cards for speedy data transfer and runs on AMD's open ROCm software stack, all inside Meta's hefty Open Rack Wide chassis.
Helios supports trillion-parameter models
Big names like TCS, Oracle, and HPE are already on board—TCS is partnering with AMD on a 200-megawatt data-center project in India, while Oracle plans to grab thousands of AMD GPUs for its own projects.
With support for trillion-parameter models and flexible Ethernet-based connections (instead of NVIDIA-only setups), Helios is built for the future of large-scale AI.