AMD Advancing AI 2026: Talking CDNA5 with AMD's Alan Smith (chipsandcheese.com)

At AMD's "Advancing AI 2026" event, Alan Smith unveiled the CDNA5 architecture, marking a pivotal shift from the older GCN architecture to a more modern RDNA-based design. The CDNA5 architecture, which powers the newly launched MI455 and Helios GPUs, focuses on enhanced efficiency and performance optimization for data center applications, especially in AI workloads. By moving to a SIMD32 structure and increasing VGPR access from 256 to 1024, AMD aims to reduce register pressure and improve throughput for AI tasks, addressing the growing demands of machine learning and AI-driven applications across its product lines. The significance of this update lies in AMD's strategic integration of AI capabilities within its GPU architecture, allowing for better resource sharing and performance across both HPC and AI workloads. This modern architecture supports high-bandwidth memory and robust cache systems designed for diverse computational tasks, enabling AMD to compete more effectively in the rapidly evolving AI landscape. With CDNA5, AMD seeks to leverage advancements in GPU technology to meet the increasing demands for AI integration in gaming, neural rendering, and more, thus influencing the future of AI-centric hardware development.