AMD’s AI Architecture Evolution: From EPYC and Instinct to Helios and Heterogeneous Integration
AMD’s data center strategy is evolving from traditional CPU- and GPU-centric chip competition toward a comprehensive AI computing platform. Across next-generation EPYC processors, Instinct GPUs, chiplets, SoIC/CoWoS advanced packaging, the Helios rack-scale architecture, custom ASICs, and CPO, a common technology direction is emerging: reducing the cost of data movement through heterogeneous integration while expanding the scope of competition from the chip level to the package, node, rack, and ultimately the entire AI cluster.
We will share the following topics
1. Custom ASICs: Extending AMD Beyond General-Purpose CPUs and GPUs
2. EPYC’s Core Advantage Is Shifting from Core Count to Chiplet Architecture
3. Memory and I/O Are Becoming the Next EPYC Bottlenecks
4. CPU and GPU Architectures Are Moving from External Connectivity Toward Platform-Level Integration
5. Helios: AMD’s Transition from a Chip Company to Rack-Scale Computing
6. Instinct GPUs: Moving from Process Scaling to Advanced Packaging Scaling
7. SoIC and CoWoS: Two Different Dimensions of GPU Scaling
8. Advanced Packaging Competition Is Shifting from Capacity Toward Yield
9. The Next Generation of Instinct: System Scaling Matters More Than Standalone GPU Performance
10. CPO: Addressing the Next Electrical I/O Bottleneck
11. AMD’s Technology Roadmap Is Fundamentally About Solving Data Movement
