By David M Williams
Publication Date: 2026-02-22 04:02:00
HPE says it will be among the first vendors to offer AMD’s Helios rack-scale AI architecture, a turnkey system designed for trillion-parameter model training and high-performance inference.
He helios design gather up to 72 AMD Instinct MI455X GPU in a single rack, delivering up to 260 TB/s of scalable bandwidth, 31 TB of HBM4 memory, and up to 2.9 exaflops of FP4 performance AI.
What distinguishes the platform is HPEThe decision to base network expansion on standards-based Ethernet rather than proprietary interconnections. The company is integrating Broadcom Tomahawk 6 silicon with Juniper networking software and Ultra Accelerator Link over Ethernet (UALoE) to create a specially designed Ethernet fabric optimized for GPU-to-GPU communication.
HPE maintains that this approach offers customers greater architectural flexibility, reduced vendor lock-in, and easier integration with…