AMD's data-centre sales nearly double as its MI450 AI chips ship

Where the growth came from

The Data Center segment, which sells server CPUs and AI accelerators, grew from roughly half its current size a year ago to US$8.5 billion. AMD names two causes in its release: strong demand for EPYC processors and the first shipments of the Instinct MI450 Series.

The rest of the company moved the other way. Client and Gaming together fell 8 per cent to US$3.7 billion. Inside it, laptop and desktop processors grew 7 per cent to US$2.9 billion, while gaming, which includes console chips, fell 40 per cent to US$773 million.

Margins and the outlook

Gross margin was 54 per cent on a GAAP basis and 56 per cent on AMD's adjusted basis. For the fourth quarter the company guided revenue to about US$17 billion, plus or minus 3 per cent.

Chief executive Lisa Su called it a record quarter and said demand for Instinct and EPYC is still strengthening. That is a company forecast; it depends on customers building the data centres they have planned.

Why it matters for engineers

A data-centre business that grows twice as fast as the company shows where chip design effort, and hiring, is going: server processors, AI accelerators and the networks and memory that feed them. Gaming chips, once a mainstay, shrank in the same quarter.

For a computer engineering student, this is the practical side of computer architecture and parallel programming: AI accelerators like the MI450 are judged by how fast they move data and how many operations they run in parallel, not by clock speed alone.

What an AI accelerator actually does

An AI accelerator such as the Instinct MI450 is built for one kind of work: multiplying very large grids of numbers, over and over. Training and running a language model is dominated by these matrix multiplications, so a chip with thousands of small arithmetic units working in parallel can finish the job far faster than a general-purpose processor with a handful of large cores.

The limit is often not arithmetic but data movement. A model's weights and activations must be fed to those units continuously, so accelerators carry stacks of high-bandwidth memory beside the processor and fast links to neighbouring chips. Engineers measure them in operations per second, memory bandwidth and energy per operation, and data-centre buyers care about the last of these because power and cooling are now a major cost.

Server CPUs such as EPYC still matter in the same racks. They run the operating system, schedule the work, handle storage and networking, and prepare data for the accelerators. AMD's quarter grew on both, which is a reminder that an AI data centre is a whole computer system, designed from the chip and memory up to the network and the cooling plant.