Research Finder
Find by Keyword
HPE Unveils AMD’s Helios Rack-Scale System as a New Open Standard for AI Factories
Solution to target cloud providers, neoclouds, and large enterprises seeking open, rapidly deployable AI infrastructure.
12/04/2025
Key Highlights:
- HPE will offer the AMD Helios rack-scale AI architecture globally in 2026, providing a fully integrated system built on the Open Compute Project’s Open Rack Wide (ORW) standard.
- The rack incorporates 72 AMD Instinct MI455X GPUs, next-generation EPYC CPUs, and 31 TB of HBM4 with approximately 1.4 PB per second of memory bandwidth.
- HPE introduced a Juniper Networking scale-up Ethernet switch co-developed with Broadcom and using the Ultra Accelerator Link over Ethernet (UALoE) standard.
- The architecture supports up to 2.9 exaFLOPS of FP4 performance per rack for large-scale model training and inference.
The News
HPE will offer the AMD Helios rack-scale AI architecture globally in 2026, delivering a fully integrated, liquid-cooled system built on the OCP Open Rack Wide standard. Each rack includes 72 AMD Instinct MI455X GPUs, next-generation EPYC CPUs, and 31 TB of HBM4 to support large-scale model training and inference. HPE also introduced a new scale-up Ethernet switch using Broadcom’s Tomahawk 6 ASIC and the UALoE standard, bringing open Ethernet to the internal fabric. The design provides high-bandwidth, low-latency connectivity suited for trillion-parameter workloads.
Analyst Take
HyperFRAME Research has been closely following the evolution of the Open Rack Wide standard, and this announcement represents one of the most complete implementations we have seen thus far. HPE’s decision to deliver Helios as a fully integrated ORW-aligned system shows how open rack-scale architectures are becoming meaningful options for organizations building out AI infrastructure at scale. Customers have been anticipating a major systems provider to take a visible leadership role around ORW, and HPE is stepping into that role at a time when demand for dense, operationally consistent AI compute continues to grow.
It is also important to acknowledge the foundational role Meta played in shaping the ORW standard through its contributions to the Open Compute Project. ORW reflects several years of Meta’s data center design experience, particularly around serviceability, power efficiency, and the realities of running high-density AI infrastructure at scale. By adopting the ORW specification, both AMD and HPE are building on work that originated in Meta’s hyperscale environment and is now becoming available to a broader market. This connection adds depth to the announcement because it represents one of the first commercially supported implementations of an open rack design that emerged from hyperscale production needs rather than vendor-driven specifications.
What stands out is the pairing of open rack design with an open scale-up Ethernet fabric. HPE is using purpose-built HPE Juniper Networking hardware and software, and collaborating with Broadcom to implement UALoE. This demonstrates that high-bandwidth, low-latency communication for AI workloads can be delivered through an Ethernet-based fabric that fits more naturally into modern data center environments. The approach aligns with organizations that want long-term architectural flexibility and a smoother path for integrating AI clusters with their existing operational models.
The Helios platform places HPE firmly within the emerging AI factory landscape. The company is offering a rack-scale system capable of supporting frontier-scale training and large inference fleets while maintaining alignment with open standards for compute, networking, cooling, and serviceability. For organizations building hybrid AI strategies or integrating model development with existing analytics platforms, this level of consistency and openness reduces deployment friction and provides a clearer foundation for long-term planning.
What Was Announced
HPE introduced a globally available offering of the AMD Helios rack-scale AI architecture, with availability in 2026, built around the Open Rack Wide standard. ORW provides a double-wide, liquid-cooled rack designed for dense compute, efficient power distribution, and streamlined serviceability. Inside this enclosure, HPE integrates AMD Instinct MI455X accelerators, next-generation EPYC processors, Pensando networking, and the ROCm software environment, creating an architecture optimized for large-scale training and inference workloads.
Complementing the Juniper Networking scale-up switch with Broadcom Tomahawk 6 chip, AMD’s Pensando Pollara 400 AI NIC (a 400 Gbps Ultra Ethernet Consortium–ready adapter now shipping to customers) delivers UEC-ready RDMA and programmable, path-aware congestion control, providing a foundational layer for the end-to-end open Ethernet fabric in Helios racks.
The system’s memory capabilities are substantial, with 31 terabytes of HBM4 across the rack and approximately 1.4 petabytes per second of aggregate memory bandwidth. HPE complements this with a purpose-built scale-up Ethernet switch developed with Broadcom, based on the Tomahawk 6 ASIC and the UALoE standard. This configuration delivers the internal bandwidth and latency characteristics needed for very large model workloads while remaining consistent with the broader Ethernet ecosystem. HPE describes Helios as a turnkey option for cloud providers, neoclouds, and large enterprises that want to deploy high-performance AI clusters while keeping their architectural choices aligned with open standards.
According to HPE, a fully configured Helios rack can deliver up to 2.9 exaFLOPS of FP4 compute performance, supported by liquid cooling and an ORW-compliant service model designed for operational sustainability. By assembling these components within a single, standards-based architecture, HPE is presenting Helios as a platform that balances performance with long-term infrastructure control.
Looking Ahead
HPE’s adoption of Helios arrives at a time when organizations are reassessing how they design and scale AI infrastructure. The market is taking a closer look at open rack-scale systems that can evolve alongside shifting workloads and vendor ecosystems. Early deployments will offer practical insight into how Helios performs under operational demands and how it fits within broader data center strategies.
Several factors will influence the next stages. Operational testing will show how effectively Helios supports large-scale training and sustained inference. The maturity of the ROCm software ecosystem, Pensando networking, and orchestration tools will shape customer confidence as clusters grow. Availability of AMD’s next-generation accelerators and CPUs will also matter, particularly for multi-rack deployments. TSMC’s confirmed CoWoS capacity ramp to 90,000 wafers per month by end-2026—more than tripling from 2024 levels—with CoWoS-L variants surging to support larger HBM4 stacks, secures the advanced packaging required for Helios to reliably deliver its full 31 TB HBM4 and 2.9 exaFLOPS FP4 performance per rack.
We will be observing how Helios transitions from announcement to early production deployments, since these real-world implementations will reveal how well the architecture performs under operational constraints. In our view, customer priorities are becoming clearer. Organizations want high-density systems that are easier to deploy, operate, and evolve, and HPE’s early adoption of ORW aligns well with those expectations.
This announcement also highlights broader industry dynamics. Open rack standards give customers more influence over how their infrastructure evolves. Open scale-up Ethernet encourages a more interoperable networking landscape and integrates more naturally across data centers. By reducing deployment complexity, open rack architectures help organizations move more quickly from experimental AI clusters to production environments with clearer operational models. And by grounding compute and networking in shared standards, the industry gains a pathway where performance and architectural flexibility can advance together.
Challenges remain, including facility readiness, power and cooling constraints, and integration work across data and management layers. Even so, organizations increasingly want AI infrastructure that is open, sustainable, and consistent with long-term architectural planning. HPE’s early commitment to ORW and UALoE places the company in a strong position as these considerations become more central to customer decision-making.
Moreover, we expect this work to provide an ecosystem boost to the Ultra Ethernet Consortium (UEC), since the HPE/AMD Helios architecture specifically uses the UALoE standard, which leverages the open, high-performance Ethernet principles defined by the UEC for AI workloads. HPE is a founding member of the UEC, and the technology described—a scale-up Ethernet switch using the Broadcom Tomahawk ASIC for low-latency, high-bandwidth communication within the rack—is an early commercial embodiment of the UEC’s core goal: adapting standard Ethernet into a fabric capable of handling massive AI and HPC traffic. As such, we anticipate that this integrated system can demonstrate the UEC’s progress in advancing an open, standards-based alternative to proprietary interconnects for trillion-parameter-scale computing.
Don Gentile | Analyst-in-Residence -- Storage & Data Resiliency
Don Gentile brings three decades of experience turning complex enterprise technologies into clear, differentiated narratives that drive competitive relevance and market leadership. He has helped shape iconic infrastructure platforms including IBM z16 and z17 mainframes, HPE ProLiant servers, and HPE GreenLake — guiding strategies that connect technology innovation with customer needs and fast-moving market dynamics.
His current focus spans flash storage, storage area networking, hyperconverged infrastructure (HCI), software-defined storage (SDS), hybrid cloud storage, Ceph/open source, cyber resiliency, and emerging models for integrating AI workloads across storage and compute. By applying deep knowledge of infrastructure technologies with proven skills in positioning, content strategy, and thought leadership, Don helps vendors sharpen their story, differentiate their offerings, and achieve stronger competitive standing across business, media, and technical audiences.
Ron Westfall | VP and Practice Leader for Infrastructure and Networking
Ron Westfall is a prominent analyst figure in technology and business transformation. Recognized as a Top 20 Analyst by AR Insights and a Tech Target contributor, his insights are featured in major media such as CNBC, Schwab Network, and NMG Media.
His expertise covers transformative fields such as Hybrid Cloud, AI Networking, Security Infrastructure, Edge Cloud Computing, Wireline/Wireless Connectivity, and 5G-IoT. Ron bridges the gap between C-suite strategic goals and the practical needs of end users and partners, driving technology ROI for leading organizations.
Share
Stephen Sopko | Analyst-in-Residence – Semiconductors & Deep Tech
Stephen Sopko is an Analyst-in-Residence specializing in semiconductors and the deep technologies powering today’s innovation ecosystem. With decades of executive experience spanning Fortune 100, government, and startups, he provides actionable insights by connecting market trends and cutting-edge technologies to business outcomes.
Stephen’s expertise in analyzing the entire buyer’s journey, from technology acquisition to implementation, was refined during his tenure as co-founder and COO of Palisade Compliance, where he helped Fortune 500 clients optimize technology investments. His ability to identify opportunities at the intersection of semiconductors, emerging technologies, and enterprise needs makes him a sought-after advisor to stakeholders navigating complex decisions.