The Conical-Fishbone Clock Tree: A Clock-Distribution Network for a Heterogeneous Chip Multiprocessor AI Chiplet

Tomas Figliolia,Andreas G. Andreou

DOI: https://doi.org/10.1109/dsd.2019.00111

2019-08-01

Abstract:The 2.5D nano-Abacus SOC is a neuromorphic CMP architecture for accelerated computing, hardware AI inference and machine intelligence. The SOC consists of a silicon interposer a 3D memory stack, a host FPGA and three (17.47 mm X 14.13 mm) computational chiplets fabricated in the Global Foundries 55nm CMOS technology. The latter employ two networks on chip, a token ring L1-NOC, and a switched circuit L2-NOC as well as a DDR DRAM PHY interface and a general purpose I/O port. In this paper we present a novel clock tree network, named Conical-Fishbone clock tree, that is employed in the low latency energy aware clock distribution networks of the chiplets. The many nets in a Fishbone clock tree consist of conical sections in an inverted cone architecture, with each resulting ring considered to be one clock tree. When a ring is excited at uniform intervals from the ring below, the symmetry in the circular characteristics of the wire makes the effect of reflections be exactly the same along any place in the wire. The designed clock tree allows low clock skew, consumes low power while offering the modularity to support a hierarchical design using standard CAD flows.