AMD held an analyst event this week. This was their second event this year. At their event in June, they announced an impressive looking GPU purpose-built for AI, the Mi300; had a number of high profile partners on stage; and gave positive commentary around a host of initiatives. And of course their stock fell the next day, apparently investors were disappointed that AMD did not actually announce any paying customers for the product nor did they release many performance metrics for the Mi300. At yesterday’s event they announced that the Mi300 matches Nvidia’s H100 performance and two major paying customers – Meta and Microsoft. So of course, the stock was down on the news.
AMD has had a great run over the past several years. They have capitalized on Intel’s stumbles and picked up a healthy amount of share in the profitable data center market. Moreover, their CEO, Lisa Su, does not get enough credit for the turnaround she put in place taking the company from a bumbler at execution, always missing deadlines, to a solid operations machine delivering compelling products at a regular clip.
Seen in that light, the latest news continues on that trend. They announced a product, got it into customers’ hands, and are now poised to generate real revenue from them. Textbook execution. The Mi300 demonstrates that they have real capabilities in the AI market, and not just in the form of a product, but an entire ecosystem of hardware partners like Dell and Lenovo, and a steadily growing roster of software partners.
To be balanced, there are still plenty of obstacles ahead. Performance is one thing, actual customer usage is another. For companies buying data center silicon, especially the hyperscalers, many factors beyond raw performance matter. Moreover, AMD’s software still has a long way to go. The big force in AI systems today is Nvidia’s Cuda software which is the de facto platform for most AI systems. AMD’s response to Cuda, called RocM, is finally a complete product but not yet a true contender for Cuda’s throne. RocM is only supported on a handful of AMD’s products, and more importantly is still largely unfamiliar to the majority of AI developers. It’s not that there is necessarily anything wrong with RocM, it’s just that Cuda has a decade+ head-start, and Nvidia is not standing still.
Which brings us to the real challenge for AMD. As much as they have been steadily soldiering along for years, the market has moved past them. Nvidia has overtaken everyone to become the leader in data center processors. In the new data center, AMD is in second place in CPUs and second place in GPUs. This has to be frustrating for a company that has worked so hard.
That being said, the company can still do well. If Nvidia really comes to dominate the data center the way Intel did for a decade with 90% share of wallet (and it is unlikely that it will), AMD is that second source of GPUs that every customer says they want as they write ever bigger checks to Nvidia. And if Nvidia’s share eventually levels off somewhere below 70%, then that means a three way contest for everything. In the past, the hyperscalers felt there was a big advantage on standardizing purchases around a small number of vendors. In the new data center, that no longer makes sense, heterogeneous compute means multiple vendors, and AMD should continue to have a seat at the table.
As much as AMD seems destined to remain the perpetual second source in the data center, that may not be such a bad place to be. They do not have to have better CPUs than Intel or better GPUs than Nvidia. So long as their GPUs are better than Intel’s and their CPUs are better than Nvidia’s CPUs, AMD will have a healthy roster of paying customers. They do not need to outrun all the bears, they just have to outrun the other racers.
Leave a Reply