HPE Launches First Turnkey AI System Powered by AMD Helios Architecture
Hewlett Packard Enterprise (HPE) has unveiled its first turnkey artificial intelligence system built on AMD's Helios rack-scale architecture. This new solution integrates advanced networking, high-performance processing, and scalable storage capabilities, positioning itself as a comprehensive platform for modern AI workloads in enterprise and research environments.
The system incorporates 72 AMD Instinct MI455X GPUs per rack, providing substantial computational resources tailored for demanding AI applications. Utilizing a scale-up Ethernet network, the platform leverages custom networking hardware and software from Juniper, alongside Broadcom's Tomahawk-6 network chip. The design follows the open UALoE (Ultra Accelerator Link over Ethernet) standard, which allows for efficient accelerator interconnects over standard Ethernet, and adheres to Open Rack Wide (ORW) specifications from the Open Compute Project. These features aim to optimize energy consumption, facilitate advanced liquid cooling, and simplify ongoing maintenance, offering a robust alternative to proprietary solutions like Nvidia's NVLink.
With its configuration, the system delivers an aggregate scale-up bandwidth of 260 terabytes per second and achieves up to 2.9 exaflops of FP4 performance. It also features 31 terabytes of HBM4 memory and a memory bandwidth reaching 1.4 petabytes per second. The inclusion of a newly developed scale-up Ethernet switch, created in collaboration with Broadcom, ensures optimized performance for AI workloads across standard Ethernet infrastructure. This switch is enhanced by HPE's AI-driven automation and quality assurance functionalities, designed to streamline network operations and reduce deployment times and associated costs.
The platform is further supported by open-source software, including AMD's ROCm suite and AMD Pensando networking technology, providing users with a flexible and extensible environment for AI development and deployment. The integration of these technologies supports the training of large-scale models involving trillions of parameters, as well as high-throughput inference tasks.
According to HPE, the system is particularly suited for organizations engaged in the development and deployment of extensive AI models, such as model providers and AI service companies. While the system is delivered as a turnkey solution, its initial price point may limit its accessibility to larger enterprises and research institutions. The architecture is also positioned as a valuable frontend solution for high-performance computing (HPC) environments, complementing existing supercomputing infrastructure by handling data-intensive AI workloads and facilitating efficient workflow integration.
This new rack-scale solution forms part of HPE's broader AI factory initiative, which envisions next-generation data centers purpose-built for artificial intelligence. These AI factories act as centralized nodes, integrating computing, storage, and networking resources to deliver scalable high-performance solutions for complex AI projects. The global interconnection of these nodes creates a unified application environment, enabling efficient resource sharing and workload management across distributed locations.
In addition to centralized AI training within these AI factories or supercomputers, HPE addresses the growing trend of moving inference closer to data sources, such as edge environments. Enhanced edge access points and the introduction of the NX 301 Multiservice Edge Router are designed to support this shift, allowing more inference operations to occur at the edge rather than in the cloud. This approach aims to reduce latency, optimize bandwidth utilization, and lower operational costs, further enhancing the efficiency of AI deployments across diverse operational landscapes.
The introduction of this AMD-based, rack-scale turnkey system by HPE reflects the continued evolution of AI infrastructure, providing organizations with scalable, high-performance solutions tailored to the rapidly expanding demands of artificial intelligence and data-intensive computing.