AMD launched new Instinct MI350 Series accelerators, previewed Instinct MI400 Series GPUs and outlined its next-gen AI rack systems that integrate the company's stack.
The upshot is that AMD is going after AI workloads as inferencing has ballooned and models for multiple use cases have proliferated.
Speaking at AMD's Advancing AI 2025 conference, CEO Lisa Su said training is critical to developing for developing models but there's a big picture. "We are seeing an explosion of models, especially models for specific uses such as coding, healthcare and finance," said Su. "Over the next few years we expect hundreds of thousands, and eventually millions of purpose-built models each tuned for specific tasks, industries or use cases."
- AMD sells ZT Systems manufacturing business to Sanmina for $3 billion
- AMD’s data center sales surge in Q1
That selection of models and use cases will drive compute requirements.
Here's what the company launched at its event:
- AMD launched its Instinct MI350 series, which has 4x more performance than its predecessor and up a 35x gain in inferencing performance. Those AI accelerators come in air-cooled and direct liquid-cooled options.
- AMD previewed its upcoming Instinct MI400 Series GPUs with a 10x performance increase.
- The next-gen Helios AI rack infrastructure was also previewed. That rack system is optimized for AI workloads and integrates M1400 GPUs, EPYC CPUs and Pensando NICs.
- AMD is also building out its software stack led by its ROCm platform. The company touted 3.5x inference gains in its upcoming ROCm 7 release.
- ROCm can run more than 1.8 million Hugging Face models.
- The company launched its AMD Developer Cloud with ROCm and AMD GPU access.
- AMD also touted traction for its Instinct GPUs with cloud providers such as AWS, DigitalOcean, Meta, Microsoft and Oracle Cloud. System giants such as Dell, HPE and Supermicro are also building out with AMD Instinct MI350 Series GPUs.
- AMD's Su laid out a roadmap through 2027 that not only includes processors but open rack-scale designs that'll be used by hardware partners. AMD recently sold the manufacturing operations of ZT Systems.
With the launches, AMD is making a play for inference workloads. Nvidia is best known for training workloads, but also has a big footprint in AI inferencing. AMD is looking to be a counterweight to Nvidia and also gain share as the AI total addressable market expands.
Indeed, AMD brought along some powerful references including Meta deployed Instinct MI300X for Llama 3 and Llama 4 inference. OpenAI CEO Sam Altman touted its AMD partnership as did Oracle, HUMAIN, Microsoft, Cohere and others.
