AMD provided the industry with the first look at its Helios rack-scale platform after it was on display at the Open Compute Project (OCP) Global Summit.

Helios is designed on Meta’s Open Rack Wide (ORW) specification and enables operators to integrate AMD’s family of hardware, including its Instinct GPUs and Pensando GPUs, with their own differentiated solutions.

Set for deployment in 2026 with OEM and ODM partners, Helios complies with standards from OCP, as well as DC-MHS, UALink, and Ultra Ethernet Consortium (UEC) – supporting both open scale-up and scale-out fabrics.

It features a double-wide layout for improved serviceability and standards-based Ethernet for multipath resiliency. There’s also quick-disconnect liquid cooling on the backside that provides efficient thermal performance at high density.

The new rack is specifically designed around AMD’s upcoming flagship GPU, the MI450 series.

At rack scale, AMD said a Helios system housing 72 of its new MI450 GPUs can provide up to 1.4 exaflops of FP8 (8-bit floating point) and 2.9 exaflops of FP4 performance, with 31-terabytes of total HBM4 memory and 1.4 Pb/s of aggregate bandwidth.

In terms of networking, the rack system is believed to feature up to 260 Tb/s of scale-up interconnect bandwidth and 43 Tb/s of Ethernet-based scale-out bandwidth.

“Open collaboration is key to scaling AI efficiently,” Forrest Norrod, EVP and GM of AMD’s data center solutions group, explained. “With Helios, we’re turning open standards into real, deployable systems – combining AMD Instinct GPUs, EPYC CPUs, and open fabrics to give the industry a flexible, high-performance platform built for the next generation of AI workloads.”

Oracle taps AMD for MI355Xs and next-gen NICs

Having recently struck a sizable hardware agreement with OpenAI, AMD continues to garner customer wins, the latest with Oracle.

The hyperscaler will be the first to deploy AMD’s Instinct MI355X GPUs for its Oracle Cloud Infrastructure (OCI) Supercluster, providing up to 2.8-times higher throughput and interconnected with ultra-low-latency RDMA networking.

The team up will also see Oracle be the first to deploy AMD’s Pensando Pollara network interface cards (NICs) in a hyperscale AI backend network. The NICs provide advanced RDMA over converged Ethernet (RoCE) functionality and programmable congestion control to clusters, meaning Oracle engineers can more efficiently manage their MI355X-powered clusters.

In addition, AMD’s 5th-generation EPYC processors are set to be used by Oracle in Oracle’s E6 instances – a deployment the pair claim will provide up to twice the performance of the prior generation at the same price point.