The Instinct MI300X, AMD’s newest AI chip, will be used, according to Meta, OpenAI, and Microsoft during a Wednesday AMD investor event. As far as technology businesses are concerned, this is the most important sign that they are looking for alternatives to the pricey Nvidia graphics processors that have been necessary for creating and launching artificial intelligence applications like OpenAI’s ChatGPT.
AMD’s newest high-end CPU may reduce the cost of creating AI models and put competitive pressure on Nvidia’s rapidly increasing sales of AI chips if it is good enough for the tech businesses and cloud service providers creating and serving AI models when it begins arriving early next year.
Lisa Su, the CEO of AMD, stated on Wednesday that “All of the interest is in big iron and big GPUs for the cloud” are the main areas of interest.
According to AMD, the MI300X is built on an innovative architecture, which frequently results in significant enhancements in performance. Its most notable feature is its 192GB of HBM3, a state-of-the-art, high-performance memory type that can accommodate larger AI models and transports data more quickly.
Su made a direct comparison between Nvidia’s flagship AI GPU, the H100, and the MI300X and the systems built with it.
According to Su, “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 come back faster, especially as responses get more complicated.”
AMD’s biggest concern is whether businesses who have been relying on Nvidia will spend the time and resources to add a different GPU provider. “It takes work to adopt AMD,” Su stated.
In order to compete with Nvidia’s industry standard CUDA software, AMD announced to investors and partners on Wednesday that it has updated its ROCm software package. This improvement handled a significant weakness that was a major factor in why AI developers currently preferred Nvidia.
Cost will also be important. While Nvidia’s chips can cost up to $40,000 per, AMD did not disclose price for the MI300X on Wednesday. Su informed reporters that AMD’s chip would need to be less expensive to acquire and run than Nvidia’s in order to entice users to purchase it.
AMD announced on Wednesday that some of the businesses most in need of GPUs had already agreed to deploy the technology. Based on a recent estimate from research firm Omidia, Meta and Microsoft were the two biggest buyers of Nvidia H100 GPUs in 2023.
According to Meta, AI inference tasks including processing AI stickers, altering images, and managing its assistant will be handled by MI300X GPUs.
According to Kevin Scott, CTO of Microsoft, the business will make MI300X chips available via its Azure online service.
The chips will also be used by Oracle’s cloud.
OpenAI announced that Triton, one of its software solutions, which is used in AI research to access chip features but isn’t as large as a language model as GPT, will support AMD GPUs.
AMD only anticipates roughly $2 billion in total data centre GPU revenue in 2024, so it isn’t expecting the chip to sell in large quantities just yet. Even if that figure includes semiconductors other than GPUs, Nvidia stated to have sold more than $14 billion worth of data centre equipment in only the most recent quarter.
AMD, however, anticipates the market for AI GPUs might grow to $400 billion in the next four years, which would double its prior estimate. This shows the high bar and growing demand for premium AI chips, which is why the company is currently concentrating investor attention on the product line.
Additionally, Su indicated to reporters that AMD does not believe defeating Nvidia is necessary for success in the market.
Regarding the AI chip industry, Su told reporters, “I think it’s clear to say that Nvidia has to be the vast majority of that right now.” “We believe it could be $400 billion-plus in 2027. And we could get a nice piece of that.”