AMD (AMD) Advancing AI Keynote summary

Market outlook and AI trends

• AI infrastructure demand is accelerating, with inference workloads expected to grow over 80% annually and total AI accelerator TAM projected to exceed $500 billion by 2028.

• The proliferation of specialized and frontier AI models is driving the need for flexible, high-performance compute across industries.

• Agentic AI is emerging as a new class of user, requiring always-on, high-performance GPUs and CPUs, effectively adding billions of virtual users to global compute infrastructure.

Product and technology roadmap

• The MI350 series was launched, delivering a 4x generational leap in AI compute, with up to 288GB memory and support for models up to 520B parameters on a single GPU.

• MI355 offers up to 40% more tokens per dollar than competitors, with leadership performance in both training and inference, and production shipments have begun.

• The MI400 series and Helios AI rack, launching in 2026, will offer up to 10x more performance for frontier models, 432GB HBM4, and 2.9 exaflops FP4 performance.

• Next-gen Venice CPUs (256 Zen 6 cores, 2nm) and Vulcano AI NIC (800Gbps) will power Helios, supporting open standards like UALink and UEC.

Open ecosystem and software innovation

• Commitment to openness spans hardware, software, and solutions, with ROCm 7 delivering 3.5x performance over ROCm 6 and supporting all major frameworks.

• ROCm Enterprise AI and the AMD Developer Cloud were announced, providing easy access to GPUs, developer tools, and industry-specific AI solutions.

• Collaboration with open-source communities (VLLM, SGLang, LLMD) is accelerating feature velocity and performance, outpacing proprietary alternatives.