In response to a question about Microsoft’s plans to compete with Google’s search business, CEO Satya Nadella indicated that he only intends to make Google ‘dance’. Looks like Microsoft sees itself in AMD, and the duo is set out to break another monopoly: that of Nvidia.   

Just a few days ago, reports emerged that Microsoft has been working on its own in-house AI processors, codenamed Athena, since 2019 in a bid to challenge Google and AWS, who have their own set of in-house chips for training and inference. To top that, new reports allege that this project, which Microsoft has been keeping under the shadows for so long, is actually being done in partnership with the chip company, AMD.

A part of Athena or not, speculations are ripe that Microsoft is financing AMD’s AI chip push. It is surprising, especially since Microsoft has a strong partnership with Nvidia, which helps to train OpenAI’s large language models on Azure. It almost seems like AMD and Microsoft have planned a coup against Nvidia. No wonder Nvidia’s shares moved lower following the report. 

It will not be the first time we’ve seen Microsoft use AMD. AMD’s silicon already powers the secure AI infrastructure in Azure cloud services, while also playing a role in the Xbox Series X and Series S consoles. 

Market positioning  

Currently, Nvidia enjoys a monopoly in the GPU market, so the industry is seeking an alternative. At least in the data centre segment, it feels like Intel and AMD are fighting over scraps, while Nvidia continues to dominate sales. It’s possible that Microsoft doesn’t want to spend so much money into Nvidia and is looking for an alternative that can handle their workload. 

“They [Microsoft] were counting on Intel, but Intel is still not able to deliver. AMD, on the other hand, has got a very good GPU technology, but they fell back in software optimisation,” says semiconductor analyst Sravan Kundojjala. AMD will be releasing the Instinct MI300 later this year which will be their first shot at building a true data centre/HPC-class APU, combining the best of AMD’s CPU and GPU technologies.

To catch up to Nvidia, AMD developed a software framework called ROCm™, an open software platform designed to provide HPC and AI communities with access to open compute languages, compilers, libraries, and tools. However, it is not yet mature enough. Kundojjala explains that Nvidia’s CUDA works with all kinds of industries, but AMD doesn’t have that luxury. In the data center, AMD is trying to focus its attention on certain verticals – certain customers, certain workloads – so it can position itself. One of those customers is Microsoft.

Nvidia alone cannot serve every use case or every customer. Kundojjala says, “it’s a rising tide, lifts all boats kind of situation.” So, if AMD wins, Nvidia doesn’t have to lose. Because the market is so big, both can grow. At the same time, he also argues that Microsoft’s AMD push will not have much to do with the pricing that Nvidia offers for its GPUs. 

“Unlike consumer applications, data centre customers are not sensitive to price. Enterprise cloud market doesn’t care about pricing. What they need is that the product has to be good, the roadmap has to be good, and the maturity of the software framework that they provide along with the hardware product – that has to be good,” he added. 

Generative AI gold rush

In the recent earnings call, AMD CEO Lisa Su suggested that there are opportunities beyond hyperscalers and game consoles for AMD’s IP, such as semi-custom opportunities with higher volume potential.

AMD has already had success with semi-custom gaming consoles such as the Sony PS and Microsoft Xbox, and it is attempting to replicate this semi-custom business model in data centers with hyperscalers like Microsoft. “In the AI space, I am pretty sure that Microsoft can help AMD accelerate its software,” said Kundojjala.

Kundojjala mentions that AMD’s data centre GPU revenue is a few hundred million dollars, which is negligible compared to Nvidia’s data centre GPU chip revenue of almost $15 to $16 billion. The reason AMD has come so late to the party is that it never really focused on it. ‘Their plate was already full; they were already eating Intel’s lunch,’ he said. 

AMD is still reporting solid profits, despite a PC inventory correction and a slowdown in data center sales, while Intel is reporting losses and investing billions in new fabs. Additionally, AMD has recently revealed its new Ryzen 7040U series processors for laptops, making bold claims that the chips not only beat the competition from Intel but also outpace the MacBook M2. Overall, they have been making significant strides in that direction. 

Now, AMD realises that generative AI is a crucial use case in data centres. However, there will not be any dramatic change since their MI300 product will ramp up until 2024. So, it will take at least another three years for AMD to make its mark in data center GPUs.