System-level Trade-offs and Optimization for Data-Driven Applications

Abstract

Data-driven applications are becoming increasingly important, fueled by the rapid rise of the Internet of Things (IoT) and Artificial Intelligence (AI). Systems must now be able to store, process and act swiftly on increasingly large amounts of data, while consuming minimum possible power. This shifts the focus to system-level integration and optimization – especially as Moore's Law slows down, and technology development at 5nm and beyond becomes increasingly harder and more expensive. SEMI has built a cross-supply-chain collaborative platform specifically to enable an early assessment of trade-offs and future technologies (5–8 years out). The first project focused on interconnect strategies, which are critical to most computing systems. We examined the performance limits for the best possible options for on-chip interconnects at technology nodes <= 20 nm. These limits highlight the need for system-level strategies, and we studied these by comparing a two-dimensional (2D) system with an interposer-based system (2.5D) to quantify the impact of the latter on the energy-delay product for various applications, especially data-driven ones.