Astera speaks softly and carries a big switch
Summary
Astera Labs has introduced Scorpio X, an AI fabric switch designed as an alternative to Nvidia's NVSwitch for building rack-scale AI systems. Unveiled on May 5, 2026, Scorpio X is an ASIC featuring 320 lanes of PCIe 6.0 connectivity and 5.12 TB/s of bidirectional bandwidth. Astera claims this large PCIe switch can enable dozens of GPUs to function as a single unit without requiring accelerator redesigns, making it compatible with nearly any accelerator. Beyond basic switching, Scorpio X incorporates in-network compute capabilities similar to NVSwitch, specifically accelerating collective communications crucial for generative AI inference, particularly with Mixture-of-Experts (MoE) models. Astera has also developed Hypercast, a multicast operation optimized for dynamic MoE inference groups. While not a direct NVSwitch competitor in raw bandwidth, Scorpio X is positioned as a vendor-agnostic solution, supporting disaggregated inference architectures and complementing Astera's expanded Scorpio P-series switches and COSMOS management suite. Production is slated for the second half of 2026.
Key takeaway
For CTOs and VPs of Engineering evaluating AI infrastructure, Astera's Scorpio X presents a compelling, vendor-agnostic alternative to proprietary interconnects like NVLink. If your strategy involves mixing and matching accelerators or deploying MoE models, Scorpio X's PCIe 6.0 connectivity and specialized in-network compute for collective communications could simplify system design and improve inference performance. You should investigate its compatibility with your existing and planned GPU deployments, especially for disaggregated inference architectures, as it offers a direct, high-bandwidth connection.
Key insights
Astera's Scorpio X offers a PCIe-based, vendor-agnostic alternative for rack-scale AI, accelerating MoE inference.
Principles
- PCIe can scale for rack-level AI fabrics.
- In-network compute boosts AI collective communications.
- MoE models benefit from optimized multicast operations.
Method
Scorpio X integrates 320 lanes of PCIe 6.0 with in-network compute and Hypercast multicast to accelerate collective communications for generative AI, especially MoE inference.
In practice
- Use PCIe switches for vendor-agnostic GPU fabrics.
- Consider Scorpio X for MoE inference acceleration.
- Implement disaggregated inference with PCIe interconnects.
Topics
- Astera Labs
- Scorpio X
- PCIe 6.0 Connectivity
- AI Fabric Switches
- Generative AI Inference
Best for: Investor, CTO, VP of Engineering/Data, AI Hardware Engineer, AI Architect, Machine Learning Engineer
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by The Register: Enterprise Technology News and Analysis.