At HPE Discover Barcelona 2025, Hewlett Packard Enterprise (NYSE: HPE) announced a major advancement in large-scale artificial intelligence infrastructure with the introduction of the industry’s first AMD “Helios” rack-scale AI architecture integrated with open, Ethernet-based scale-up networking. The new solution is designed to help cloud service providers, including emerging neocloud operators, deploy AI training and inference environments faster and with greater efficiency.
The rack-scale system combines AMD’s “Helios” AI compute platform with purpose-built HPE Juniper Networking hardware and software, powered by Broadcom’s latest Tomahawk 6 silicon. Built on the open Ultra Accelerator Link over Ethernet (UALoE) standard, the architecture is engineered to support extremely large AI models, high-throughput inference workloads, and next-generation training requirements involving trillions of parameters. The solution will be delivered by HPE Services teams with deep expertise in liquid-cooled infrastructure and exascale-class deployments, enabling customers to accelerate implementation while reducing operational risk.
Antonio Neri, President and Chief Executive Officer of HPE, said the collaboration builds on more than a decade of joint innovation with AMD in high-performance computing. He noted that the new rack-scale platform and networking capabilities give cloud providers a more flexible and lower-risk path to expanding AI capacity, while benefiting from open standards and faster time to deployment.
Dr Lisa Su, Chair and CEO of AMD, highlighted that the “Helios” platform extends the long-standing partnership with HPE by bringing together AMD’s full AI compute stack with HPE’s system-level engineering. She said the result is an open, rack-scale solution that delivers higher efficiency, scalability, and performance for AI-driven workloads.
The AMD “Helios” rack-scale system is designed in alignment with Open Compute Project specifications and supports advanced power delivery, direct liquid cooling, and serviceability for next-generation AI environments. Each rack can connect up to 72 AMD Instinct™ MI455X GPUs, delivering up to 260 terabytes per second of scale-up bandwidth and as much as 2.9 AI exaflops of FP4 performance. The platform also provides 31 terabytes of fourth-generation high-bandwidth memory and 1.4 petabytes per second of memory bandwidth to support data-intensive AI and HPC applications.
HPE is also introducing a standards-based Ethernet scale-up switch and software optimized specifically for the “Helios” architecture. Developed in collaboration with Broadcom, this solution is the first to deliver AI-optimized scale-up performance over standard Ethernet. Integrated automation and assurance capabilities simplify network operations, reduce deployment complexity, and lower overall costs, while complementing HPE’s existing scale-out and scale-across AI networking offerings.
Broadcom’s President and CEO, Hock E. Tan, said the collaboration reflects a shared commitment to open Ethernet innovation. He added that Broadcom’s high-performance networking silicon enables ultra-low latency, lossless operation, and scalable performance required by modern AI data centers, while giving customers greater flexibility and choice.
HPE (NYSE: HPE) is a global provider of enterprise technology, integrating artificial intelligence, cloud platforms, and networking solutions to help organizations drive better outcomes. Through continuous innovation and deep technical expertise, HPE supports the evolution of how businesses operate and how people work. The company enables customers across industries to improve efficiency, gain actionable insights from data, and extend their reach and influence.
