AMD pushes for open ecosystem to challenge Cuda dominance

AMD is hoping to win over artificial intelligence (AI) developers in Asia-Pacific with its open software ecosystem and help the region compete on a global stage without relying on proprietary AI development tools.

While the US chipmaker has made hardware strides with its Instinct datacentre graphics processing units (GPUs) and Ryzen processors, AMD’s software layer – specifically the ROCm platform – is critical to unlocking that performance for the broader market.

But rather than build a walled garden where the technology is controlled by a single entity, AMD has gone down the open source route with ROCm, which integrates with other open source projects such as vLLM to allow for faster innovation.

And while doing so, it hopes to break the grip of closed-source software environments, particularly Nvidia’s Cuda, that are used to build AI applications that run on GPUs.

“We could try to build something that’s closed source, but we won’t get the velocity of an open ecosystem,” said Anush Elangovan, AMD’s vice-president of AI software. “Instead, we want to leverage everyone’s capabilities to move the industry forward. It’s like the Linux kernel, where everyone collaborates and moves really fast.”

Elangovan noted that the open nature of AMD’s software stack is particularly beneficial for markets in Asia-Pacific, as it allows for a “common baseline” that lets companies build up their AI capabilities.

“I’ve been involved with quite a few companies in the region that are building large datacentres and deploying AMD chips at scale,” he said, adding that ROCm will also allow them to compete in both model development and infrastructure.

The company has also embarked on its “ROCm everywhere” initiative to unify the experience for developers, whether they are working on a laptop or a massive supercomputer. This allows students and startups in the region to start small with affordable hardware and scale up as needed.

This scalability is supported by AMD’s chiplet architecture, which Elangovan claimed provides distinct advantages in inference workloads, the process of running live data through a trained model.

He noted that AMD’s chips are designed with high memory bandwidth, allowing them to handle very large AI models on a single system. This efficiency can reduce the need for liquid cooling systems, which are expensive to retrofit in older facilities.

"You can go for a little less density, so you can do air-cooled infrastructure versus liquid-cooled infrastructure, and then still get the capabilities that are top of the line," he explained.

In terms of workloads, Elangovan said besides large language models (LLMs), organisations are also running more text-to-image and text-to-video workloads. He cited Luma Labs as an example, noting that their Ray3 video generation model is “fully trained and serving on AMD platforms.”

Even as AMD continues to iterate on ROCm – with version 7 supporting the new MI350 chips – Elangovan said developers should not view AMD solely as a hardware supplier.

“AMD is increasingly building and shipping software as a software company,” he said. “You should think of us as a software platform that developers can trust and build on, one that will outlive generations of hardware.”