Ualink 1.0 is official: new open technology connects up to 1,024 gpus and challenges Nvidia’s domain

The Ualink consortium officialized the final specification of technology UALink 1.0establishing a new open interconnection standard for artificial intelligence and high performance computing accelerators (HPC).

With support for up to 1,024 GPUS or other accelerators in the same systemthe initiative aims to compete directly with the NVLinkproprietary nvidia technology widely used in clusters such as Blackwell NVL72.

Created by giants like AMD, Intel, Google, Apple and Broadcom, Ualink wants to democratize the construction of AI superstructures by eliminating dependence on exclusive – and expensive – solutions from Nvidia.

Focus Speed ​​and Efficiency

Ualink 1.0 offers a bidirectional transfer rate of up to 200 gigatransfers per second (GT/s) By range, with signs at 212.5 GT/s to accommodate error and coding corrections. The connections can be configured in x1, x2 or x4, reaching until 800 gt/s by link in four -track versions.

Each Ualink -based system can scales horizontally with up to 1,024 accelerators connected by switcheseach with a unique 10 -bit identifier to ensure precise routing and deterministic performance in up to four racks.

Low latency, high connectivity

The connections here were optimized for cables up to 4 meters, delivering round trip latency less than 1 microsecondeven with 64b or 640b loads. The protocol stack is divided into four optimized hardware layers: physical, data, transaction and protocol layer.

The physical layer is based on Ethernet patterns (such as 200gbase-KR1/CR1), while the link layer adds data in 640 bytes units, applying integrity verification (CRC) and optional repetition logic.

Direct transactions and operations between memories

The transaction layer draws attention for having compressed address and efficiency of up to 95% in the use of the protocol during real loads. It also supports direct operations between memories – reading, writing and atomic transactions – maintaining order between local and remote memory spaces.

With this level of integration, Ualink truly positions itself as a robust solution for data centers that require scalability, performance and interoperability between different manufacturers.

Embedded security and confidential computing support

Ualink also brings advanced security features, such as the UALinkSecwhich implements encryption and authentication directly on the hardware. This guarantees protection against physical violation and allows the use of Reliable Execution Environments (TEE)like AMD SEV, ARM CCA and Intel TDX – fundamental for the so -called confidential computing.

In addition, architecture allows the concept of “Virtual Pod Partitioning”where different groups of accelerator operate in isolation within the same physical system, enabling multi-tenant loads in shared infrastructures.

Modern Management and Integration via APIs

Management of Ualink Pods will be done through firmware agents and control software integrated with standard interfaces as PCIe e Ethernet. The system brings complete support to APIS REST, telemetry, workload control and failure insulation, meeting the requirements of modern operations on a large scale.

With the arrival of 200g Ualink 1.0 specification, consortium companies are building an open ecosystem for scale accelerator interconnection. We are excited to see the solutions that will emerge in the market and how they will boost the next generation of applications in IA

Peter OnufrykPresident of the Ualink Consortium

Also read:

Ualink versus nvidia: a new dispute on the horizon

Nvidia continues to lead the industry with solutions such as NVLINK and Racks Blackwell NVL72, which connect up to 72 GPUS by rack and climb to 576 GPUS in a POD. By 2026, the company plans to launch the Vera Rubin platform with a support for up to 144 GPUS per rack, while the expected Rubin Ultra for 2027 will increase this capacity to 576 GPUs per rack.

Ualink, however, represents a significant change in allowing the construction of similar architectures using standardized and affordable componentswhich can boost new competitors and reduce dependence on a single supplier.

Source: Ualink