Particle.news
Download on the App Store

HPE Adopts AMD’s Helios AI Racks for 2026 Launch With JuniperBroadcom Scale‑Up Switch

The design leans on open Ethernet-based interconnects to counter vendor lock-in.

Overview

  • HPE will offer AMD’s Helios rack-scale architecture globally in 2026, integrating a purpose-built Juniper scale-up switch developed with Broadcom using Ultra Accelerator Link over Ethernet.
  • Each 72‑GPU rack built on the OCP Open Rack Wide form factor targets up to 2.9 exaFLOPS of FP4 performance with Instinct MI455X accelerators and next‑gen EPYC “Venice” CPUs, presenting the rack as a single large GPU pod.
  • The HPE Juniper switch is reported to use Broadcom Tomahawk 6 silicon delivering about 102.4 Tbps aggregate bandwidth to provide high-bandwidth, low-latency connectivity over standard Ethernet.
  • HLRS selected an HPE Cray GX5000 system named Herder using AMD Instinct MI430X GPUs and EPYC “Venice” CPUs, with delivery planned for the second half of 2027 and service targeted by year-end 2027.
  • HPE is tying Helios into its AI-native and hybrid-cloud portfolio, including GreenLake and converged ArubaJuniper networking, with primary customers expected to be cloud service providers and neoclouds.