Advanced Micro Devices Inc. needs to flawlessly execute the launch of its first graphics chips targeted to the AI market, amid big expectations for the second half of the year, which now seem hinged specifically on the fourth quarter.
Even as AMD
AMD,
forecast better-than-expected revenue for its third quarter Tuesday, Wall Street was more fixated on the fourth quarter, when the company is now set to launch its first graphics processor (GPU) chips to accelerate and run artificial intelligence on servers in data centers.
The obsession is understandable, especially since rival Nvidia Corp.
NVDA,
doesn’t need to wait until the end of the year to cash in on this explosive opportunity. Nvidia told investors last quarter that it expects a major boost from its own graphics accelerator chips in the July period, one factor that could fuel record quarterly revenue for the chip company.
AMD said it plans to launch its so-called MI300 family of chips in the fourth quarter, and it is currently sampling them with its leading AI, high-performance-computing and cloud customers. Chris Rolland, an analyst at Susquehanna Financial, said in an email he had been hoping for the launch in September.
The pressure the company is feeling from customers and Wall Street to get its chips out fast and flawlessly was palpable on its call with analysts after the earnings. Many analysts, instead of focusing on the ongoing quarter, leaped to the fourth quarter and beyond, to get a sense of how much revenue AMD expects from MI300.
“We are expecting a large ramp in the second half for our data-center business and weighted toward the fourth quarter,” Chief Executive Lisa Su told analysts. She still expects roughly 50% growth in the second half of the year relative to the first half.
When another analyst asked if she was expecting GPU sales of around $500 million this year, Su cautioned that she did not want to go into that level of granularity.
“I think your number may be a little bit high in terms of the GPU sales,” she replied, but reiterated a point from Chief Financial Officer Jean Hu, who projected high-single-digit growth in the company’s data-center business on a year-over-year basis.
AMD’s initial projected revenue for the MI300 is expected from a supercomputer developed by Hewlett Packard Enterprise Inc.
HPE,
called El Capitan. Other customers outside the world of high-performance supercomputers appear to be clamoring for these chips as well, to speed up compute power required for handling AI training and machine learning, but sales to other customers will likely occur in the first half of 2024.
“As we go into 2024…the customer interest on MI300 X is very high,” Su said. “There are a number of customers that are looking to deploy as quickly as possible, so we would expect early deployment as we go into the first half of 2024, and then we would expect more volume in the second half of 2024 as those things fully qualify.”
In his preview note, Rolland suggested that the bar on AMD had been set too high.
Bernstein Research analyst Stacy Rasgon also reminded investors that he had heard of “lofty expectations for the business in 2024 and beyond.” But because AMD is just sampling their AI GPUs in Q3, “real volume might not be until the 2H (second half) of next year, which might make it prudent to temper some of those expectations,” Rasgon said in his preview note to clients last week.
Investors seemed satisfied — or relieved— enough by the responses from AMD’s management, with shares of the company up nearly 3% in after-hours trading. The stock is up about 81% so far this year.
But AMD will need to execute well in order to capture the opportunities ahead of it. While the company has gotten used to racking up wins recently, it will have to avoid the sorts of mishaps with execution that it experienced under past leaders — the kinds of issues that rival Intel Corp.
INTC,
has experienced more recently with its long-delayed Sapphire Rapids chip.
AMD certainly isn’t alone in this market, facing big competition from Nvidia, eventually Intel and startups like Cerebras Systems, among others. As customers clamor to get their own AI initiatives off the ground, they could easily be prepared to move to whichever designer that has the most available chips.