Large chip vs. MCM for a high-performance system

The capabilities of modern semiconductor processes enable us to design a system on a chip. This implies a large function on a large chip. However, as we approach-and eventually exceed-a 1-GHz clock rate, it is valid to ask whether partitioning this chip into a few smaller chips and using a multichip module (MCM) to interconnect the functions is better for performance and cost. This article explores the design, performance, and cost issues for a large chip and an MCM, focusing primarily on the on-chip and off-chip delays and the trade-offs between the two. It also shows performance trends for each approach, to help us understand future performance limits for microprocessors.