AI giveth and AI taketh CPU

In this episode of The Stack Overflow Podcast, host Ryan Donovan sits down with Mark Papermaster, CTO of AMD, to explore how AMD is leading the charge in AI-driven computing through innovative chip design, open ecosystems, and strategic partnerships. Papermaster outlines AMD's decade-long focus on heterogeneous computing, combining CPUs and GPUs on the same chip or chiplet architecture to deliver high performance and energy efficiency. He emphasizes the company's use of chiplets—modular, partitioned silicon designs—to improve manufacturing yields, reduce bottlenecks, and enable rapid adaptation to evolving AI workloads like inference and agentic flows. The discussion highlights AMD’s open software stack, ROCm, which enables flexibility and collaboration across diverse hardware and software environments. Papermaster also reveals how AI is now accelerating AMD’s own chip design process through agentic workflows, unlocking performance gains that human engineers alone couldn’t achieve. The episode concludes with forward-looking insights on the future of AI, including the rise of small language models at the edge, the importance of energy efficiency across the entire compute stack, and AMD’s expanding role in hyperscale AI infrastructure through partnerships with OpenAI and Meta. Key takeaways include: AMD’s chiplet-based architecture enables scalability, efficiency, and faster time-to-market; open software (ROCm) and hardware standards (like Open Rack) reduce vendor lock-in and drive industry-wide innovation; AI is now being used to design better chips, creating a virtuous cycle of performance gains; energy efficiency is achieved through a holistic approach spanning transistors, packaging, software, and data center power management; and the future of AI will be defined by tailored, efficient solutions for specific workloads rather than one-size-fits-all models.