HPE introduces AMD ‘Helios’ rack-scale architecture to support large-scale AI deployments

Hewlett Packard Enterprise (HPE) has announced the availability of what it describes as the industry’s first AMD “Helios” AI rack-scale architecture with integrated scale-up Ethernet networking. According to HPE, the offering is aimed at helping cloud service providers (CSPs), including emerging neocloud operators, accelerate large-scale AI training and inference deployments.

The solution combines AMD’s “Helios” AI rack-scale architecture with HPE Juniper Networking hardware and software, and Broadcom’s Tomahawk 6 networking chip. HPE stated that the architecture is based on the Ultra Accelerator Link over Ethernet (UALoE) open standard and is designed to support the networking demands of large AI models, including trillion-parameter training workloads and high-throughput inference.

According to the company, the solution will be delivered by HPE Services teams with experience in liquid-cooled infrastructure and large-scale, exascale-class installations. HPE said the focus is on providing customers with flexibility, interoperability, improved energy efficiency and faster deployment timelines as demand for AI compute capacity continues to rise.

“For more than a decade, HPE and AMD have worked together on high-performance computing systems, including multiple exascale-class deployments,” said Antonio Neri, President and CEO of HPE. Neri said the introduction of the AMD “Helios” architecture, combined with HPE’s scale-up networking, is intended to help cloud service providers deploy and scale AI infrastructure with greater flexibility and reduced operational risk.

AMD highlighted the collaboration as an extension of its long-standing partnership with HPE. “With ‘Helios’, we are bringing together AMD’s compute technologies and HPE’s system design capabilities to deliver an open, rack-scale AI platform,” said Dr Lisa Su, Chair and CEO of AMD. She added that the platform is designed to improve efficiency and scalability for AI workloads.

Rack-scale AI architecture and networking approach

HPE said it will be among the first vendors to offer the AMD “Helios” rack-scale solution, which is based on specifications from the Open Compute Project (OCP). According to the company, these specifications are optimised for power delivery, liquid cooling and serviceability, requirements that are increasingly critical for next-generation AI systems.

As outlined by HPE, the rack-scale system connects up to 72 AMD Instinct MI455X GPUs per rack. The company claims the configuration can deliver up to 260 terabytes per second of aggregated scale-up bandwidth and up to 2.9 AI exaflops of FP4 performance. The system also includes 31 terabytes of high-bandwidth memory (HBM4) and up to 1.4 petabytes per second of memory bandwidth, which HPE says is intended to support large AI and high-performance computing workloads.

A key element of the offering is what HPE describes as an industry-first standards-based Ethernet scale-up switch and software platform developed specifically for the AMD “Helios” architecture. The company said the switch has been designed to optimise AI workload performance over standard Ethernet and has been developed in collaboration with Broadcom.

According to HPE, the networking platform incorporates AI-native automation and assurance capabilities from HPE Juniper Networking, with the goal of simplifying network operations and reducing deployment time and operational costs. The scale-up solution is positioned as part of HPE’s broader “networks for AI” portfolio, which also includes scale-out and scale-across networking options.

Focus on open standards and interoperability

HPE emphasised that the AMD “Helios” rack-scale architecture is built on open standards, including OCP’s Open Rack Wide (ORW) specifications. The double-wide rack design is intended to support energy-efficient liquid cooling while remaining serviceable for ongoing operations.

The architecture also uses AMD’s ROCm open-source software stack and AMD Pensando networking technology. According to the companies, this approach is designed to encourage innovation, reduce vendor lock-in and lower total cost of ownership. HPE added that the Ethernet-based HPE Juniper Networking switch, built on open and established communication standards, is intended to enable faster feature updates and greater ecosystem flexibility.

Broadcom said it is working closely with HPE on the development of the scale-up Ethernet switch for the AMD “Helios” architecture, building on the companies’ existing collaboration around Ethernet-based data centre networking.

“Together with HPE and AMD, we are working to advance open, Ethernet-based AI infrastructure for scale-up environments,” said Hock E. Tan, President and CEO of Broadcom. He added that Broadcom’s networking silicon is designed to support low-latency, high-performance and scalable AI workloads, while giving customers more choice through standard Ethernet-based architectures.