A Serial H-Tree Router for Two-Dimensional Arrays

Sam Fok, Kwabena Boahen

Abstract: Existing routing mechanisms for two-dimensional (2D) arrays either use low-overhead grids with one or two shared wires per row or column (e.g., RAM) or high-overhead meshes with many wires connecting neighboring clients (e.g., supercomputers). Neither is suitable for intermediate-complexity clients (e.g., small clusters of silicon neurons). We present a router tailored to 2D arrays of such clients. It uses a tree laid out in a fractal pattern (H-tree), which requires less wiring per signal than a grid, and adopts serial-signaling, which keeps link-width constant, regardless of payload size. To route from the tree's leaves to its root (or vise versa), each node prepends (consumes) a delay-insensitive 1-of-4 code that signals the route's previous (next) branch; additional codes carry payload. We employ this serial H-tree router to service a 16x16 array of silicon-neuron clusters, each with 16 spike-generating analog somas, 4 spike-consuming analog synapses, and one 128-bit SRAM. Fabricated in a 28-nm CMOS process, the router communicates 26.8M soma-generated and 18.3M synapse-targeted spikes per second while occupying 43% of the client's 35.1x36.1um^2.