If AMD's latest high-end chip is good enough for the technology companies and cloud service providers building and serving AI models when it starts shipping early next year, it could lower costs for developing AI models and put competitive pressure on Nvidia's surging AI chip sales growth.
AMD CEO Lisa Su said all of the interest is in big iron and big GPUs for the cloud and the MI300X is based on a new architecture, which often leads to significant performance gains. Its most distinctive feature is that it has 192GB of a cutting-edge, high-performance type of memory known as HBM3, which transfers data faster and can fit larger AI models.
Su directly compared the MI300X, and the systems built with it to Nvidia's main AI GPU, the H100.
"What this performance does is it just directly translates into a better user experience. When you ask a model something, you'd like it to return faster, especially as responses get more complicated."
The main question facing AMD is whether companies building on Nvidia will invest the time and money to add another GPU supplier. "It takes work to adopt AMD," Su said.
AMD told investors and partners that it had improved its software suite called ROCm to compete with Nvidia's industry-standard CUDA software, addressing a key shortcoming that had been one of the primary reasons AI developers prefer Nvidia.
Price will also be important. AMD didn't reveal pricing for the MI300X, but Nvidia's can cost around $40,000 for one chip, and Su told reporters that AMD's chip would have to cost less to purchase and operate than Nvidia's to persuade customers to buy.
AMD said it had already signed up some of the companies, including Meta and Microsoft. Meta said it will use MI300X GPUs for AI inference workloads such as processing AI stickers, image editing, and operating its assistant.
Microsoft's CTO, Kevin Scott, said the company would offer access to MI300X chips through its Azure web service. Oracle's cloud will use the chips. OpenAI said it would support AMD GPUs in one of its software products, called Triton, which isn't a big language model like GPT but is used in AI research to access chip features.