Reiner Pope (MatX): Designing AI Chips From First Principles for LLMs

Reiner Pope, co-founder and CEO of MatX, discusses the company's mission to create advanced chips for large language models (LLMs). Pope details his and his co-founder's departure from Google to address the economic challenges of running increasingly large and expensive models like GPT-3. He addresses initial investor skepticism about competing with NVIDIA and the importance of software lock-in, explaining how the balance between hardware and software costs has shifted, making multi-platform solutions more viable. MatX's chip design focuses on high matrix-multiplied performance and a hybrid SRAM-HBM memory system to optimize both latency and throughput. Pope also touches on the company's use of agentic AI in chip design and the challenges of scaling manufacturing to meet the growing demand for data center compute power.