Chip firm AMD has unveiled a new Instinct accelerator, the MI325X.

Announced by the company’s CEO Lisa Su at this year’s Computex event in Taipei, Taiwan, the AMD Instinct MI325X accelerator is expected to be available to customers from Q4 2024.

Featuring 288GB of HBM3E memory and 6TBps peak memory bandwidth, eight accelerators will fit into the Instinct MI325X Platform, which is built using the same architecture as the MI300X server platform.

“MI325X extends our leadership in generative AI,” Su said during her Computex keynote speech, adding that an advanced server running eight MI325X accelerators can run advanced models up to 1 trillion, which she said is double the size supported by an H200 server.

AMD said that when compared to “the competition” – again, Nvidia GPUs – the MI325X boasts a 2x memory capacity, 1.3x bandwidth, and 1.3x better AI compute performance.

In the small print, AMD claims that this metric has been reached by comparing the projected specifications of the MI325X accelerator, which is expected to have 288GB HBM3e memory capacity and 1.3 petaflops of FP16 and 2.6 petaflops of FP8 peak compute performance, with a selection of Nvidia GPUs.

AMD said that the highest published results for the Nvidia Hopper H200 141GB, Blackwell HGX B100 192GB, and Blackwell HGX B200 192GB all demonstrate a lower memory capacity and bandwidth performance than the MI325X.

The MI325X will run on the company’s CDNA 3 architecture, which has specifically been designed for data center applications.

AMD updates its data center GPU roadmap

In addition to announcing the MI325X, Su also used her Computex keynote to show off AMD’s updated roadmap, which will see the company release a new data center GPU every year, starting with the MI325X.

“We have now expanded our product roadmap so it's on an annual cadence, which means a new product every year,” Su said during the event.

The company also unveiled its incoming CDNA 4 architecture set to be released in 2025, which will be used to power the AMD Instinct MI350 series accelerators and deliver up to a 35x increase in AI inference performance compared to CDNA 3.

Due for release in 2025, the AMD Instinct MI350x accelerator will be the first product to use this new architecture. Containing 288GB of HBM3e, the MI350x will be manufactured using a 3nm process and will support FP4 and FP6 data types.

A new CDNA architecture has also been planned for release in 2026 to power the AMD Instinct MI400 series accelerators.

Earlier this year, after the company posted its Q1 2024 financial results, Su labeled the Instinct MI300 AI chips the “fastest-ramping product in AMD history,” noting that it had passed $1 billion in total sales in less than two quarters.

Based on the company’s “expanding customer engagements,” Su added that the company expected data center GPU revenue to exceed $4bn in 2024.

Last month, Microsoft announced it would be making AMD’s MI300X accelerators available to customers through the company’s Azure cloud computing service, with Scott Guthrie, EVP of Microsoft’s Cloud and AI group, describing the MI300X as the “most cost-effective GPU out there right now for Azure OpenAI.”