UC San Diego
Cross-Layer Pathfinding for Off-Chip Interconnects
- Author(s): Srinivas, Vaishnav
- Advisor(s): Kahng, Andrew B
- et al.
Off-chip interconnects for integrated circuits (ICs) today induce a diverse design space, spanning many different applications that require transmission of data at various bandwidths, latencies and link lengths. Off-chip interconnect design solutions are also variously sensitive to system performance, power and cost metrics, while also having a strong impact on these metrics. The costs associated with off-chip interconnects include die area, package (PKG) and printed circuit board (PCB) area, technology and bill of materials (BOM). Choices made regarding off-chip interconnects are fundamental to product definition, architecture, design implementation and technology enablement. Given their cross-layer impact, it is imperative that a cross-layer approach be employed to architect and analyze off-chip interconnects up front, so that a top-down design flow can comprehend the cross-layer impacts and correctly assess the system performance, power and cost tradeoffs for off-chip interconnects. Chip architects are not exposed to all the tradeoffs at the physical and circuit implementation or technology layers, and often lack the tools to accurately assess off-chip interconnects. Furthermore, the collaterals needed for a detailed analysis are often lacking when the chip is architected; these include circuit design and layout, PKG and PCB layout, and physical floorplan and implementation.
To address the need for a framework that enables architects to assess the system-level impact of off-chip interconnects, this thesis presents power-area-timing (PAT) models for off-chip interconnects, optimization and planning tools with the appropriate abstraction using these PAT models, and die/PKG/PCB co-design methods that help expose the off-chip interconnect cross-layer metrics to the die/PKG/PCB design flows. Together, these models, tools and methods enable cross-layer optimization that allows for a top-down definition and exploration of the design space and helps converge on the correct off-chip interconnect implementation and technology choice. The tools presented cover off-chip memory interfaces for mobile and server products, silicon photonic interfaces, 2.5D silicon interposers and 3D through-silicon vias (TSVs). The goal of the cross-layer framework is to assess the key metrics of the interconnect (such as timing, latency, active/idle/sleep power, and area/cost) at an appropriate level of abstraction by being able to do this across layers of the design flow. In additional to signal interconnect, this thesis also explores the need for such cross-layer pathfinding for power distribution networks (PDN), where the system-on-chip (SoC) floorplan and pinmap must be optimized before the collateral layouts for PDN analysis are ready.
Altogether, the developed cross-layer pathfinding methodology for off-chip interconnects enables more rapid and thorough exploration of a vast design space of off-chip parallel and serial links, inter-die and inter-chiplet links and silicon photonics. Such exploration will pave the way for off-chip interconnect technology enablement that is optimized for system needs. The basis of the framework can be extended to cover other interconnect technology as well, since it fundamentally relates to system-level metrics that are common to all off-chip interconnects.