AMD's Long and Winding Road to the Hybrid CPU-GPU Instinct MI300A

AMD's journey to develop the hybrid CPU-GPU Instinct MI300A began in 2012 when the US Department of Energy invested in AMD for research in memory technologies and hybrid computing. Over a decade later, this investment has led to the creation of the powerful "El Capitan" supercomputer using the Instinct MI300A compute engine. Key researchers recently published a paper detailing AMD's evolution from concept to reality. The company received significant funding from programs like FastForward and DesignForward, totaling over $637.8 million, to develop exascale hardware designs. AMD transitioned from monolithic designs to chiplet designs and explored technologies like processor in memory and NVRAM. The MI300A architecture tightly couples CPU and GPU complexes, offering unified code execution similar to CPU-only architectures. The company faces decisions on whether to offer MI300A servers or discrete CPU-GPU pairings based on market demand for compute ratios, price, and performance. While comparisons with Nvidia's offerings are pending, AMD's advancements showcase a significant milestone in creating a true APU and a complex compute engine socket.