AMD launches MI325X AI chip, taking on Nvidia’s Blackwell

AMD has unveiled its latest artificial intelligence chip, the Instinct MI325X, to compete directly with Nvidia’s data center graphics processors.

Announced during a recent event, AMD stated that production of the MI325X will commence by the end of 2024, marking a significant step in the company's efforts to capture a larger share of the burgeoning AI chip market.

As the demand for advanced generative AI technologies like OpenAI’s ChatGPT continues to surge, the necessity for powerful GPUs in data centers has intensified.

AMD’s CEO, Lisa Su, noted that “AI demand has continued to take off and exceed expectations,” signaling a robust investment climate in AI across various sectors.

Currently, Nvidia is dominant in the data center GPU market, commanding approximately 90% of it.

Still, AMD is determined to increase its presence in what is projected to be a $500 billion market by 2028.

Although specific details about cloud or internet customers for the MI325X were not disclosed, AMD has previously indicated that tech giants like Meta and Microsoft utilize its AI GPUs, with OpenAI also reportedly using them for select applications.

The pricing for the new chip remains under wraps, as AMD typically sells these GPUs as part of complete server solutions.

To better compete with Nvidia, AMD plans to accelerate its product rollout schedule, with annual chip releases to capitalize on the AI boom.

The MI325X follows the MI300X, which began shipping late last year, and future chips, including the MI350 for 2025 and MI400 for 2026, are already in the pipeline.

The introduction of the MI325X will put it in direct competition with Nvidia’s upcoming Blackwell chips, expected to ship in significant quantities early next year.

A successful launch could attract investor interest in AMD, which has seen its stock rise only 20% this year, in stark contrast to Nvidia's impressive 175% surge.

Despite the optimism surrounding the new AI chip, AMD faces challenges due to Nvidia’s established CUDA programming language, which locks many developers into its ecosystem.

To counter this, AMD is enhancing its ROCm software to facilitate smoother transitions for developers looking to migrate their AI models to AMD's accelerators.

AMD’s AI accelerators are particularly touted for applications involving content creation and predictive analytics, thanks in part to their advanced memory technology.

Su claimed that the MI325X can deliver up to 40% more inference performance than Nvidia's H200 when running Meta’s Llama AI model.

While AI accelerators are capturing significant attention, AMD’s core business remains in central processors (CPUs), crucial for nearly every server globally.

The company reported that its data center sales exceeded $2.8 billion in the June quarter, doubling year-over-year, although AI chips represented only about $1 billion of that figure.

AMD commands about 34% of total spending on data center CPUs, still trailing behind Intel’s Xeon line.

To challenge this dominance, AMD introduced its new EPYC 5th Gen CPUs during the same event.

The new lineup features various configurations, from an economical 8-core chip priced at $527 to a 192-core, 500-watt processor designed for supercomputing at $14,813 per unit.

Su emphasized that “today’s AI is really about CPU capability,” underscoring the importance of CPUs in powering AI workloads and data analytics applications.