OCP Global Summit 2025 | Astera Labs | Scaling AI with PCIe, Ethernet, and UALink Retimers

Introduction

At this year's OCP Global Summit, Astera Labs' talk, "Scaling AI with PCIe, Ethernet, and UALink Retimers," was not just a product launch but a manifesto on the architectural evolution of AI Infrastructure 2.0.

As AI model sizes grow explosively, Astera stressed:

"The server is no longer the unit of compute; the entire rack is the new compute unit."

Presented by Anchul Sharma and Chris Blackburn, the talk laid out how Astera is building an open interconnect foundation for future AI architectures, spanning chip design, interconnect protocols and the software management ecosystem.

Content

1. Astera's Mission: Intelligent Connectivity for AI

Astera opened by defining its positioning:

"We don't just make general-purpose connectivity chips; we are an intelligent connectivity platform purpose-built for AI workloads."

Its products span three layers:

1. ICs / Boards / Modules: including retimers, gearboxes, I/O chiplets and AI fabric switches;

2. Smart Cable Modules: signal-conditioning circuitry integrated into active electrical cables (AECs), reducing system latency and cabling complexity;

3. Cosmos software platform: Astera's software-defined management layer, providing rack-wide diagnostics, monitoring and automated operations.

Cosmos is seen as Astera's "AI fabric brain," able to control and optimize link behavior uniformly across multi-protocol environments (PCIe, CXL, Ethernet, UALink).

2. AI Infrastructure 2.0: The Rack Is the New Unit of Compute

Astera drew a clear dividing line:

Architecture	Definition	Characteristics
AI Infrastructure 1.0	The server is the basic unit	4–8 GPUs per server, connected over Ethernet
AI Infrastructure 2.0	The rack is the basic unit	Hundreds of GPUs interconnected by a high-bandwidth, low-latency fabric

This shift means traditional scale-out Ethernet connectivity is no longer enough to support training of large AI models;

a scale-up fabric (such as UALink, NVLink or CXL interconnect) designed specifically for GPU-to-GPU communication must be introduced.

In Astera's view, in this new era:

"A rack is no longer a collection of servers; it is treated as a single supercomputer."

Its goal is to build an open, standardized rack-scale architecture so OEMs and hyperscalers can integrate components from different vendors in a modular way.

3. Four Pillars: Astera's Rack-Scale Strategy

Astera's architectural vision revolves around four core principles:

1. Open Standards

   • Covering PCIe, CXL, Ethernet and UALink.

   • Advocating open collaboration rather than closed ecosystems.

2. Intelligent Software

   • Cosmos as a cross-protocol management platform, providing diagnostics, monitoring and automatic configuration.

3. Multi-Protocol Flexibility

   • Freely switching among PCIe / CXL / Ethernet / UALink within or between racks depending on the application.

4. Purpose-Built Portfolio

   • Aries series (PCIe retimer / gearbox)

   • Taurus series (Ethernet retimer)

   • Leo series (CXL memory controller)

   • Scorpio series (AI fabric switch)

4. Core Products and Their Roles

(1) Aries Smart PCIe Retimer / Gearbox

• Already deployed in volume on mainstream AI platforms at PCIe Gen4/Gen5;

• The new Gen6 retimer has entered high-volume production, with performance and reliability ahead of competitors;

• The gearbox version bridges Gen6 and Gen5 systems, solving speed mismatches.

(2) Taurus Smart DSP Retimer (Ethernet)

• Speeds from 100G → 800G;

• Supports integration into active cables and modules;

• Designed for AI scale-out networks (such as data center interconnect).

(3) Leo CXL Smart Memory Controller

• Provides memory pooling and expansion;

• Lets GPUs access disaggregated memory more efficiently.

(4) Scorpio AI Fabric Switch (P / X Series)

• P series: handles mixed traffic (GPU ↔ SSD, NIC);

• X series: optimized for high-density GPU fabrics, supporting scale-up applications.

• Astera's fastest-growing product line and a key component for UALink implementations.

5. Open Rack Reference Platform

In the exhibit area, Astera deployed a reference architecture called Open Rack:

• XPU tray: integrating GPUs, CPUs, SSDs and NICs;

• PCIe switch board: using the Scorpio P series to connect devices;

• Scale-up switch: using the Scorpio X series to build a high-bandwidth GPU fabric;

• Retimers and gearboxes placed between the motherboard and backplane to ensure signal integrity.

Astera noted that this design is not a single-brand solution but an extensible open blueprint,

on which any vendor can base its own rack designs.

6. Modularity and Serviceability by Design

To close the talk, Astera highlighted an important trend: Server Modularity & Serviceability.

• Retimers, switches and cable modules will exist as standalone modules;

• They can be assembled as card-edge, mezzanine, backplane or near-chip as needed;

• Making rack repair and upgrades faster while reducing supply chain risk.

Astera's strategy is to make every key interconnect component "pluggable" and "standardizable,"

further driving the formation of an open AI Infrastructure 2.0.

Conclusion

Astera sent a clear signal at OCP 2025:

The core of AI Infrastructure 2.0 is not the number of GPUs, but the intelligence and openness of the interconnect.

Astera has evolved from a "retimer chip supplier" into an "AI fabric platform company,"

and through coordinated integration of PCIe, Ethernet, CXL and UALink,

it is leading the industry toward composable, standardized, open AI rack architectures.

Further Perspectives

1. Technical implications

   • Astera is building a "cross-protocol interconnect integration layer," something like a future AI fabric BIOS.

   • The Cosmos platform represents the software-ization of hardware-layer interconnect and is an early form of rack-level automation.

2. Supply chain observations

   • Astera's open-module strategy will reshape the supply chain structure for retimer, switch and cable modules.

   • Through collaboration with the UALink, OCP and CXL consortia, it has steadily embedded itself in the mainstream AI server ecosystem.

3. Market trends

   • AI Infrastructure 2.0 will be the industry's central theme for the next three years.

   • The "Rack = Compute Unit" concept will drive a new wave of AI SuperRack standardization.