AMD unveils Helios rackscale AI infrastructure and 6th Gen EPYC processors for enterprise data centers

AMD has introduced its Helios rackscale AI infrastructure, a comprehensive system designed to integrate its latest hardware, including 6th Gen EPYC "Venice" processors and Instinct MI455X GPUs. The Helios platform is built to support large-scale AI workloads, featuring 72 MI455X accelerators and 31 terabytes of HBM4 memory per rack. The system utilizes AMD's Pensando networking and Infinity Fabric to achieve 260 terabytes per second of scale-up bandwidth. Alongside the hardware, AMD highlighted its software advancements, specifically the ROCm Hyperloom and ROCm.AI, which are designed to optimize AI model performance and reduce inference costs. The company also showcased its Kria AI robotics platform, which provides a turnkey solution for autonomous robotics, integrating brain, spine, joint, and sensor components. These developments represent a strategic shift for AMD, moving beyond isolated components to offer complete, integrated machine solutions for enterprise and AI applications. The company is currently shipping these solutions to customers, with a ramp-up expected through the second half of the year.

Tech4K StaffVideo report

AMD has introduced its Helios rackscale AI infrastructure, a comprehensive system designed to integrate its latest hardware, including 6th Gen EPYC "Venice" processors and Instinct MI455X GPUs. The Helios platform is built to support large-scale AI workloads, featuring 72 MI455X accelerators and 31 terabytes of HBM4 memory per rack. The system utilizes AMD's Pensando networking and Infinity Fabric to achieve 260 terabytes per second of scale-up bandwidth. Alongside the hardware, AMD highlighted its software advancements, specifically the ROCm Hyperloom and ROCm.AI, which are designed to optimize AI model performance and reduce inference costs. The company also showcased its Kria AI robotics platform, which provides a turnkey solution for autonomous robotics, integrating brain, spine, joint, and sensor components. These developments represent a strategic shift for AMD, moving beyond isolated components to offer complete, integrated machine solutions for enterprise and AI applications. The company is currently shipping these solutions to customers, with a ramp-up expected through the second half of the year.

The Helios rackscale system integrates 72 Instinct MI455X GPUs and 31 terabytes of HBM4 memory for high-performance AI workloads. AMD's 6th Gen EPYC "Venice" server processors, based on the Zen 6 architecture, offer up to 512 threads per socket.

The Kria AI robotics platform provides a fully integrated solution for autonomous robotics, covering brain, spine, joint, and sensor functions. ROCm Hyperloom and ROCm.AI software tools are designed to autonomously optimize AI model performance and reduce inference costs.

AMD is leveraging Pensando networking and Infinity Fabric to deliver 260 terabytes per second of scale-up bandwidth in the Helios system. The company is shifting its strategy to provide complete, integrated machine solutions rather than focusing solely on individual components.

Chapter guide

AMD Advancing AI 2026 OverviewAn introduction to the AMD Advancing AI 2026 event, highlighting new AI hardware and robotics solutions.Foundation Robotics DemoA demonstration of the Foundation humanoid robot, which is powered by AMD's Kria AI and Ryzen Embedded X100 series.Keynote HighlightsA summary of the key product launches, including the 6th Gen EPYC processors and the Instinct MI455X GPU.Helios Rackscale ArchitectureA detailed look at the Helios rackscale system, its components, and its performance capabilities.AMD Robotics Developer PlatformAn overview of the AMD Robotics Developer Platform and its partner ecosystem.Luma AI IntegrationA demonstration of Luma AI's video generation capabilities running on AMD hardware.Instinct MI350P BenchmarksA look at the performance of the Instinct MI350P GPU in real-world benchmarks.ROCm Hyperloom and SoftwareA discussion of AMD's software stack, including ROCm Hyperloom and its role in optimizing AI workloads.Supermicro FlexTwin RackA detailed look at the Supermicro FlexTwin rack, which features 42 nodes and high-density compute capabilities.

Worth noting

  • Performance claims are based on AMD's internal testing and third-party benchmarks, which may not reflect real-world performance in all environments.
  • Availability of the Helios rackscale system and 6th Gen EPYC processors is subject to ramp-up schedules and customer demand.
  • The demonstration of the Foundation humanoid robot and Luma AI integration are specific use cases and may not represent the full capabilities of the underlying platforms.

Watch the original video ↗