Eliyan Sponsor Talk

Title: Universal Memory Interface (UMI): A practical solution to break the memory wall for AI ASICs

Abstract:

Compute performance demand has been growing exponentially in recent years, and with the advent of Generative AI, this demand is growing even faster. Moore’s law coming to an end as well as the Memory Wall (memory bandwidth & capacity) and IO Wall are the main performance bottlenecks. The chiplet system-in-package (SiP) is the industry’s solution to these bottlenecks. Advanced substrates with Silicon bridges, such as CoWoS-L or EMIB, are industry’s main solution to connect chiplets in SiPs, but they introduce several new bottlenecks. The yield of such advanced substrates drops exponentially as the substrate area size grows. For example, as the substrate area increase by 2x, its effective yield will be Y^2, where Y is the substrate yield at 1x area. The largest advanced substrate going to production is ~3000mm^2, which is smaller than 1/3 the largest standard package substrate. Thus, SiPs with advanced substrates have limited compute & memory chiplets, thus limited performance. This presentation introduces Universal Memory Interface (UMI), a high bandwidth efficient D2D connectivity technology between XPU-Memory, which help reduce the complexity and size of the advanced substrates thus enable practical implementation of large SiP for AI applications.

Speaker:

Ramin Farjadrad

Co-Founder and CEO, Eliyan

Session Link