AMD Helios Rack Claims Over 18,000 CDNA 5 GPU Compute Units and 31TB of HBM4 Memory

A single rack. More than 18,000 GPU compute units. That’s the headline claim AMD posted to X on 26 September 2026, describing its Helios rack-scale AI platform in numbers that are difficult to picture in everyday terms.

AMD states that one Helios rack contains 72 Instinct MI455X GPUs built on the company’s CDNA 5 data-centre architecture, 4,600 Zen 6 CPU cores drawn from its 6th-generation EPYC “Venice” processors, and 31 terabytes of HBM4 high-bandwidth memory. The system also incorporates AMD Pensando networking and runs on the ROCm software ecosystem. According to AMD’s official product material, the aggregate HBM bandwidth reaches 1.7 petabytes per second — meaning the system can, in theory, move more than a petabyte of data every second between memory and compute.

What AMD Is Actually Claiming

The compute figures are striking on their face. AMD says a complete Helios rack delivers up to 2.9 exaflops of FP4 compute and 1.4 exaflops of FP8 compute. Each MI455X GPU carries up to 432GB of HBM4 memory, which is how the rack accumulates 31TB across 72 cards. Scale-up bandwidth — the speed at which GPUs talk to each other within the rack — is quoted at 260 terabytes per second, with scale-out bandwidth to other racks at 43 terabytes per second.

These are vendor-supplied specifications. AMD’s figures have not been independently audited, and real-world performance will depend on workload type, software stack, power availability, cooling configuration, and how the system is actually deployed. That’s not a knock on AMD specifically — it applies to every company publishing peak-performance numbers for AI hardware.

The Architecture Behind the Numbers

CDNA is AMD’s dedicated data-centre GPU architecture, separate from the RDNA line used in consumer graphics cards. Zen 6 is the CPU architecture underpinning the EPYC Venice processors in the rack. HBM4 — high-bandwidth memory in its fourth generation — is designed to keep AI accelerators fed with data during both training and inference, where memory bandwidth is often the binding constraint rather than raw compute.

AMD describes Helios as designed for frontier-model training, fine-tuning, large-scale inference, and what it calls agentic AI workloads. Agentic AI refers to systems capable of carrying out multi-step tasks, using tools, or interacting with other software with limited human intervention — a category that has attracted growing interest from AI developers over the past two years.

Dr Lisa Su, AMD’s chair and chief executive, presented Helios at the company’s Advancing AI 2026 event in San Francisco on 22 and 23 July 2026. AMD says the system is now in production and intended for deployment by major AI companies at gigawatt scale, though independently verified shipment volumes were not available at the time of writing.

Where Helios Sits in the Market

Helios is positioned by AMD as a direct rival to large-scale AI infrastructure from Nvidia and other hardware providers. But claims of superiority over competing systems were not independently verified from the sources available. The broader industry trend is clear enough: integrated rack-scale platforms are being built because the power, memory, networking, and cooling demands of the largest AI models have outgrown what conventional server configurations can practically deliver.

AMD previewed Helios in 2025 before the full rack-scale announcement at Advancing AI 2026. The company is betting that combining compute, memory, networking, and software in one engineered system — rather than leaving customers to assemble components from multiple vendors — will appeal to hyperscale operators building AI infrastructure at the largest scales.

And the numbers do reflect that ambition. Over 18,000 GPU compute units in a single rack is not a figure that has much precedent in commercially announced systems. Whether those units translate into the claimed exaflop figures under real workloads is a question that independent benchmarks will eventually answer.

What This Means for Kent Residents

No Helios installation, planning application, or deployment involving Kent has been announced, and no local business or public-sector organisation was identified as directly connected to this system. For UK consumers and organisations more broadly, the relevance is longer-term: as AI computing capacity expands, services built on infrastructure like Helios — from NHS diagnostic tools to research computing — may eventually become more capable or more affordable. Any large data-centre deployment in the UK to house systems of this scale would also carry significant demands on electricity networks, fibre connectivity, and planning approvals, all of which are live considerations for regional authorities including those in Kent.