HPE, AMD, Broadcom Collaborate on Open AI Rack-Scale Architecture
Event summary
- HPE unveiled the 'Helios' AI rack-scale architecture, a joint effort with AMD and Broadcom, targeting cloud service providers.
- The solution utilizes AMD Instinct MI455X GPUs (72 per rack) and Broadcom's Tomahawk 6 networking chip, supporting trillion-parameter model training.
- The architecture incorporates open Ultra Accelerator Link over Ethernet (UALoE) standards and is slated for worldwide availability in 2026.
- HPE is leveraging HPE Juniper Networking hardware and software for the scale-up Ethernet networking component.
- The system delivers 260 terabytes/second of aggregated bandwidth and up to 2.9 AI exaflops of FP4 performance.
The big picture
HPE's Helios architecture represents a strategic shift towards open, rack-scale AI infrastructure, addressing the growing demand for AI compute capacity among cloud providers. By partnering with AMD and Broadcom, HPE aims to reduce vendor lock-in and accelerate deployment cycles, positioning itself as a key player in the burgeoning AI data center market. This move underscores the increasing importance of specialized hardware and networking solutions to meet the demands of increasingly complex AI models.
What we're watching
- Market Adoption
- The success of Helios hinges on CSP adoption; early deployments and customer feedback will be crucial indicators of its viability beyond initial trials.
- Open Standards
- The reliance on open standards like UALoE and OCP specifications could accelerate innovation but also introduces dependencies on external contributions and potential fragmentation.
- Competitive Response
- Other infrastructure vendors will likely respond to Helios with competing solutions, potentially intensifying price pressure and requiring HPE to continually innovate.
