AMD Helios: its first AI rack with 72 Instinct MI455X GPUs and 31 TB of HBM reaches 2.9 EXAFLOPS

Building on the announcement of the expanded collaboration between AMD and Microsoft Azure to deliver a suite of AI solutions including CPU, GPU, networking and software systems, AMD has officially unveiled the details of its Helios Rack on which the entire deployment will be based.

It is about the AMD’s first rack-scale AI system, Designed to compete directly against the Vera Rubin architecture and NVIDIA’s NVL72 system, This new equipment for data centers will debut, among other environments, in Microsoft’s Azure infrastructure.

With this launch, AMD seeks to take market share from NVIDIA, which currently dominates more than 95% of the data center GPU sector. The company’s strategy seeks to compete with the green company by offering lower total cost of ownership per token and based on the flexibility of the open standards.

Each AMD Helios includes 72 AMD Instinct MI455X GPUs and 31TB of HBM capable of reaching 2.9 EXAFLOPS

The Helios rack is made up of 18 trays or modules. Each of them houses four Instinct MI455X accelerators under the new architecture CDNA 5 and a 6th generation EPYC CPU, known by the code name “Venice”.

In total, therefore, we have 72 AMD Instinct MI455X GPUs capable of reaching 2.9 EXAFLOPS of FP4 power1.4 EXAFLOPS FP8 and access to a total of 31 TB of HBM4 memory with 19.6 TB/s of bandwidth for each GPU.

The platform uses UALink over Ethernet for “scale-up” connectivity (internal connectivity) between GPUs, reaching 260 TB/s of added bandwidth inside the rack, and 43 TB/s of “scale-out” (external connectivity) through networks AMD Thinking “Vulcan” with 800 Gbps connectivity. AMD highlights that this design offers More than 50% more memory bandwidth than the competition.

AMD bets on open standards: UALink vs NVIDIA NVLink

One of the main differences of Helios compared to its direct competition from NVIDIA is its architecture based on open networking standards. While NVIDIA’s NVL72 uses the proprietary NVLink interconnect, AMD uses UALink (Ultra Accelerator Link) for the internal interconnection between the GPUs in the rack. For external connectivity to other racks (scale-out), use the specifications of Ethernet from the Ultra Ethernet Consortium (UEC). All hardware is integrated under the standard OCP Open Rack Wide.

The platform includes liquid cooling, vertical power distribution, and a modular design that makes it easy to repair and replace without rewiring.

AMD plans to ship the first engineering samples during the second half of 2026, scheduling the start of mass production for the second quarter of 2027.