Industry Articles

A SignatureIP technical white paper on scaling cache coherency from small clusters to large heterogeneous SoCs

SignatureIP Engineering Team
Center of Excellence for Compute and I/O Subsystem IP

Abstract

Cache coherency at small agent counts is solved by broadcast snooping, which works well up to roughly eight coherent agents but does not scale. Snoop traffic grows as the square of agent count, response bandwidth grows linearly, and aggressive broadcast variants reach a practical ceiling around 128 cores. Every production SoC at modern scale is directory-based. The design problem is how to scale the directory itself: sparse and coarse-vector storage formats, direct-transfer features (DMT, DCT, DWT in AMBA CHI Rev E.b), and home-node distribution are the three mechanisms that let the same protocol scale from 8 to 256 agents. SignatureIP's C-NOC is the production-grade implementation, with a UCIe die-to-die interface in the current product for the chiplet path; multi-die coherency extensions are on the roadmap. The portfolio handles the system problems that come with scale: NC-NOC, Proxy Cache, ATC, AXI2CHI, CHI2AXI, PCIe Controller, CXL Controller, CXL Subsystem, AXI4-LLL, Ethernet IP, and the Inoculator topology generator.