AMD has rolled out its Helios AI rack, a liquid-cooled monster designed to pick a fight with Nvidia’s AI hardware empire.
The platform is aimed at frontier AI and sovereign computing, with Instinct MI455X GPUs, sixth-generation EPYC CPUs, and Pensando networking chips crammed into a single full-stack rack.
AMD first teased Helios at Advancing AI 2025, then showed early platforms at OCP 2025. The rack has now turned up on the Computex show floor looking like something built to upset expensive procurement departments.
The plan is not subtle. AMD wants Helios to challenge Nvidia’s Oberon and upcoming Kyber racks with open standards, rack-level integration and less vendor lock-in than Jensen’s leather-jacketed money printer.
The heart of the system is the Instinct MI400 series, based on AMD’s CDNA 5 architecture. The MI455X powers Helios, while the MI450X targets training and inference and the MI430X aims at HPC and sovereign AI.
The MI455X delivers 40 PFLOPs of FP4 and 20 PFLOPs of FP8 compute. AMD reckons that is double the MI350 series, while Nvidia’s Rubin GPU offers 50 PFLOPs of FP4 and 17.5 PFLOPs of FP8.
Memory is where AMD is making its louder noise. The MI455X carries 432GB of HBM4, up 50 per cent from 288GB of HBM3e, with 19.6TB/s of bandwidth.
Nvidia’s Rubin GPU comes with 288GB of HBM4 at 22TB/s, so AMD is pushing capacity while Nvidia leans harder on bandwidth. For large models and fat context windows, that extra memory might matter.
Helios uses AMD’s sixth-generation EPYC Venice CPUs, based on Zen 6. AMD says the chips are its first HPC product in volume production on TSMC’s 2nm process.
TSMC’s 2nm process moves from FinFET to nanosheet transistors. It promises 10 to 15 per cent more performance at the same power, 25 to 30 per cent lower power at the same performance and up to 15 per cent better transistor density.
The Venice chips have been teased with up to 256 cores and 512 threads. AMD claims more than 70 per cent better performance and efficiency, plus more than 30 per cent higher thread density.
Networking is handled by Pensando Vulcano 800 AI NICs and Salina DPUs. AMD claims Vulcano offers 800Gbps Ethernet throughput and up to 2.4Tbps of scale-out bandwidth per GPU.
The rack itself uses Meta’s Open Rack Wide standard, which was submitted to the Open Compute Project. It has 18 compute trays, six switches, 72 MI455X GPUs and one EPYC Venice CPU per tray.
Each Helios rack is fully liquid-cooled, weighs about 5,000 pounds, and reportedly costs $5 million to $5.5 million. It chews through roughly 225kW to 245kW of power, which should keep the local electricity board grinning.
AMD says Helios reaches 2.9 exaflops of FP4 compute, 1.4 exaflops of FP8 compute, 31 TB of HBM4 memory, 260 TB/s of scale-up bandwidth, and 43 TB/s of scale-out bandwidth. Microsoft, OpenAI, Meta, Oracle, HPE, TCS, Celestica, Nutanix and the US Department of Energy are named customers, while ROCm support covers PyTorch, TensorFlow, JAX, Hugging Face, vLLM and other AI plumbing.
