Astera Labs: Scaling AI With PCIe, Ethernet, And UALink Retimers

Author: Anchul Sharma & Chris Blackburn, Astera Labs Collection: 2025 OCP Global Summit - Expo Hall Sessions

Summary

Astera Labs presents its vision for AI Infrastructure 2.0 where the rack replaces the server as the unit of compute, with clusters of 32, 64, 128+ GPUs operating as a single high-performance system. The company's portfolio spans four protocol families: PCIe (Aries retimers, Gen 4-6, shipping millions of units to hyperscalers), CXL (Leo smart memory controllers for memory expansion and pooling), Ethernet (Taurus smart DSP retimers from 100G to 800G), and UALink/NVLink (Scorpio AI fabric switches). Astera Labs positions itself as the foundation for open rack-scale architectures, contrasting with proprietary closed ecosystems.

Key products include: Aries PCIe retimers (leading vendor for Gen 6, shipping in high-volume production while competitors are at early evaluation), the new Aries PCIe smart gearbox (bridging Gen 6 to Gen 5 data rate mismatches without throughput loss, available in 8-to-16 and dual 8-to-16 lane configurations), Taurus Ethernet retimers (100G-800G), Leo CXL memory controllers (expansion and pooling), and Scorpio AI fabric switches (P series for mixed head-node traffic between GPU/switch/SSD/NIC, X series for high-bandwidth backend scale-up). All products are managed through Cosmos (Connectivity System Management and Optimization Software) providing fleet diagnostics and management. The company has 7+ R&D centers globally with recent openings in Vietnam, Singapore, Irvine, and Seattle. Demonstrations at booth B33 include the open rack reference architecture, PCIe 6 fabric, 200G/lane Ethernet cabling, CXL memory chatbot inference, and OpenBMC rack management.