UALink IP
UALink IP refers to semiconductor intellectual property (IP) implementing the Ultra Accelerator Link (UALink) specification, an open die-to-die and accelerator interconnect standard designed for large-scale AI infrastructure and accelerator clusters.
UALink IP enables semiconductor vendors and SoC designers to integrate UALink-compatible connectivity into AI accelerators, chiplets, switches, retimers, and networking devices. The technology targets high-bandwidth, low-latency communication between accelerators within AI training and inference systems.
The UALink ecosystem is supported by the UALink Consortium, whose founding members include AMD, Intel, Google, Meta, Microsoft, AWS, Alibaba, Apple, Cisco, Astera Labs, Hewlett Packard Enterprise, and Synopsys.
Overview
UALink was developed to address scalability limitations in AI accelerator interconnect architectures. The specification defines a high-speed interconnect fabric enabling direct communication between accelerators across large AI compute clusters.
The first generation, known as UALink 200 (UALink_200), is based on a 200 Gb/s PHY derived from IEEE P802.3dj technology. The architecture is designed to support coherent and shared-memory communication between accelerators inside AI pods containing hundreds or thousands of devices.
UALink systems typically include:
- AI accelerators or AI chiplets
- UALink switch fabrics
- High-speed PHY and SerDes interfaces
- Shared-memory transport mechanisms
- Routing and fabric management logic
The standard is intended for scale-up AI infrastructure rather than general-purpose networking.
Architecture
A UALink deployment generally consists of hosts, accelerators, and UALink switches interconnected in a scalable fabric.
Each accelerator receives a unique routing identifier within the fabric. UALink switches can interconnect up to 1024 accelerators inside a single AI pod architecture.
The protocol stack separates the physical layer from higher-level accelerator communication protocols. Accelerators may connect to host processors using technologies such as:
- PCI Express
- Compute Express Link (CXL)
- AMD Infinity Fabric
- CHI C2C
- XGMI
UALink primarily focuses on accelerator-to-accelerator communication and scalable shared-memory connectivity.
Types of UALink IP:
UALink IP may include several categories of semiconductor IP blocks:
- UALink Controller IP: Implements protocol management, packet handling, routing, flow control, and link initialization functions required by the UALink specification.
- UALink PHY IP: Provides the high-speed physical interface layer, including serializer/deserializer (SerDes) technology, equalization, clock recovery, and signal integrity functions.
- UALink Switch IP: Implements switching fabrics enabling scalable accelerator interconnect topologies for AI clusters and AI pods.
- UALink Verification IP: Verification IP (VIP) used for protocol validation, compliance testing, simulation, and interoperability testing during chip development.
- UALink Chiplet IP: Chiplet-oriented implementations designed for die-to-die integration and heterogeneous accelerator architectures.
Applications
UALink IP is primarily targeted at:
- AI training clusters
- Large language model (LLM) infrastructure
- High-performance computing (HPC)
- AI inference servers
- Hyperscale cloud infrastructure
- Accelerator chiplets
- Scale-up AI systems
The technology is particularly relevant for systems requiring high-bandwidth shared-memory communication between many accelerators.
Industry Significance
UALink emerged as part of a broader industry effort to create open interconnect standards for AI infrastructure, reducing dependence on proprietary accelerator fabrics.
The specification competes and coexists with other interconnect technologies used in AI systems, including:
- NVLink
- Compute Express Link
- PCI Express
- Ultra Ethernet
The consortium model is intended to encourage interoperability between accelerators and infrastructure components from different vendors.
History
The UALink Consortium was formally established in 2024 to promote and standardize the technology.
The first public specification, UALink 1.0 (UALink_200), became publicly available in 2025 through an evaluation download program.
The initial specification focused on 200 Gb/s connectivity and AI scale-up architectures. Future versions are expected to increase bandwidth and expand interoperability capabilities for next-generation AI infrastructure.
Related Articles
- How Ultra Ethernet And UALink Enable High-Performance, Scalable AI Networks
- Reimagining AI Infrastructure: The Power of Converged Back-end Networks
Related Products
- UALink PHY + Controller
- Simulation VIP for UALink
- UALink Subsystem
- UALink Controller
- UALink PCS IP Core
See all 12 related products in the Catalog
Related Blogs
- UALink Under the Hood: Why Full-Stack Verification Wins
- Securing Scale-Up AI: Cadence’s Complete UALink Solution
- Securing UALink: Introducing Synopsys UALinkSec_200 Security Module
- Validating UPLI Protocol Across Topologies with Cadence UALink VIP
- Verification of UALink (UAL) and Ultra Ethernet (UEC) Protocols for Scalable HPC/AI Networks using Synopsys VIP
Related News
- INNOSILICON Unveils Full Suite of High-Speed UALink IP, Powering Next-Gen Token Factories with Enhanced Productivity
- Credo Introduces Industry’s First 224G Multiprotocol AI Scale-Up Retimer Supporting UALink, ESUN and Ethernet
- Arteris Joins UALink Consortium to Accelerate High-Performance AI Networks Scale Up
- Marvell Expands Custom Compute Platform with UALink Scale-up Solution for AI Accelerated Infrastructure
- Ultra Accelerator Link Consortium (UALink) Welcomes Alibaba, Apple and Synopsys to Board of Directors
The Pulse
- How Cache Coherency Simplifies AI Software
- Multiple neuromorphic chips, one grid, no new middleware
- Hardware-managed heterogeneous high-bandwidth memory and flash in LLM inference systems
- LACE: Large Language Model Aided Multi-Agent Framework for Agile RISC-V Instruction Extension
- Beyond Trusted: What the NSA’s New Guidance Means for Hardware Security
- First Benchmarks Revealed for Jalapeño, OpenAI’s Clean-Sheet General Purpose AI Accelerator ASIC
- A scalable, shader-programmable vector graphics GPU core for low-power MCUs
- MIPS Leads Open Standards Development for Physical AI Platforms as RISC-V International Premier Member
- SEALSQ’s Subsidiary IC'Alps Advances QASIC Roadmap for Sovereign, Quantum-Resistant Application-Specific Integrated Circuits
- NVMe 2.4 Update Adds Post-Quantum Security, Power Controls
- Agentrys Raises $24.5 Million to Build Agentic Design Automation for Chipmakers
- Navigating ISO 26262 Part 11: A Guide for Semiconductor Architects
- Achronix Announces Availability of High-Performance JESD204C IP Core for Speedster7t FPGAs
- Embedded Security explained: Digital signatures
- Apple’s processor cadence: A pending stutter-step for improved long-term edge inference?