Meeting the Explosive Demands of AI with Chiplet Architectures

Abstract

The era of generative AI is accompanied by an unprecedented increase in compute, memory, and bandwidth requirements of AI workloads. Chiplet architectures and advanced packaging offer a promising path to meeting these demands through much more tightly integrated compute and memory units, enabling higher compute densities and bandwidths with lower latency and power. Heterogeneous integration coupled with open interface standards can also enable targeted architectures to accelerate specific use cases through co-packaging of diverse chiplets from different sources. I will discuss how IBM’s deep investments in AI and advanced packaging can lead to new opportunities through the emerging chiplet ecosystem.