Intel, AMD and tech companies to collaborate on alternative to Nvidia NVLink

A consortium of tech companies, including AMD, Intel and HPE, will collaborate on an alternative to Nvidia's NVLink interconnects. The new standard is called Ultra Accelerator Link and should accelerate data transfer between GPUs in AI networks.

The partnership consists of AMD, Broadcom, Cisco, Google, Hewlett Packard Enterprise, Intel, Meta and Microsoft. The companies will work on a standard that allows data to be exchanged between GPUs in data centers, specifically GPUs intended for AI applications. Ultra Accelerator Link, or UALink, should become a counterpart to Nvidia's NVLink, which is now widely used for those applications. The technology enables direct data transfer between memory modules, which allows data to be transferred much faster than, for example, with PCIe connections.

Many specifications about the standard are not yet known. The companies expect to have a 1.0 version of the specification in the third quarter of this year. Around that time, the first companies can also start using UALink. The companies say that this specification should make it possible to connect 1024 accelerators within a cluster.

The companies that are now working together all build their own parts of AI infrastructure, such as servers and chipsets. The accelerators that companies like AMD are currently making do not yet work well with the hardware of other companies, while Nvidia has its own ecosystem with NVLink and its own infrastructure. By developing their own standard, the companies hope to be able to compete better with Nvidia, which leaves the competition far behind in the field of AI.