AMD Launches New GPUs for AI. The Question Is Whether They’ll Be Able to Surpass Nvidia’s Offerings

Nvidia seems to be making significant strides in the artificial intelligence sector. Its chips are highly sought after, but the competition is gradually catching up, as evidenced by AMD’s latest developments.

Meet the MI325X. On Thursday, AMD unveiled the Instinct MI325X lineup, its new accelerator graphics cards for data centers. The chips, successors to the MI300X, are built on the CDNA 3 architecture, and according to the company, they’re up to 40% more powerful in AI inference than Nvidia’s B200.

Memory improvements. The new GPUs are equipped with 256 GB of HBM3E memory, which enables impressive bandwidths of up to 6 TB/s. AMD claims that this provides 1.8 times the capacity offered by Nvidia’s H200 and 1.3 times bandwidth.

Unknown price. AMD hasn’t yet disclosed the pricing for its new chips, which are typically sold as part of a complete server. However, the company announced that servers from Dell, Eviden, Gigabyte, HP, Lenovo, and Supermicro integrating these specialized graphics cards will be launched in the first quarter of 2025.

Blackwell needs to watch out. In March, Nvidia presented the spectacular B200 chips with the new Blackwell architecture, but they still won’t be available for purchase for a few weeks. As such, this seems to be a good time for AMD to present its new models. Production of the MI325X will begin before the end of the year, which means AMD’s latest GPUs will offer a highly competitive proposal compared to Nvidia’s solutions.

A promising roadmap. In the second half of 2025, AMD will release the MI350 lineup, which will be based on the company’s CDNA 4 architecture. The upcoming GPUs are designed to deliver a colossal leap in inference performance: up to 35 times more than 2023’s MI300X accelerators. By 2026, AMD is expected to release the MI400, but details are still unknown.

AMD wants a slice of the pie. Nvidia has dominated the data-center GPU market in the last few years. Analysts estimate that the company holds up to 90% of this market, allowing it to enjoy 75% gross margins on its products.

However, AMD CEO Lisa Su is confident that the demand for data-center GPUs will continue to grow, and AMD is determined to claim its share of the market. In fact, the company projects that this segment will be worth $500 billion by 2028.

CUDA is a major challenge. The issue lies not just with the chips themselves but also with the widespread use of CUDA, Nvidia’s proprietary software platform, which holds a dominant position in the market and is closely linked to its chips. In 2016, AMD introduced its own software, ROCm, in an effort to provide a viable alternative for developers. Similarly, Intel is also working towards the same objective.

Image | AMD

Related | OpenAI Will Reportedly Lose $44 Billion by 2028 and Become Profitable a Year Later