Coral NPU (github.com)

Google Research has open-sourced Coral NPU, a compact neural processing unit IP aimed at ultra-low-power SoCs for wearables like hearables, AR glasses and smartwatches. Designed as a 32-bit RISC‑V AI accelerator, Coral NPU combines matrix, vector (SIMD) and scalar processor elements to deliver on-device inference without cloud dependence. Making the IP freely available lowers the barrier for custom silicon vendors and researchers to integrate dedicated ML acceleration into constrained, battery-sensitive products and helps expand the RISC‑V ecosystem in edge AI. Technically, Coral NPU implements an RV32IMF core with the Zve32x vector extension plus CSR/fence and bit-manipulation support (rv32imf_zve32x_zicsr_zifencei_zbb). It uses a four-stage pipeline with in-order dispatch and out-of-order retire, four-way scalar and two-way vector dispatch, 128‑bit SIMD lanes (256‑bit planned), a 32‑bit address space, 8 KB ITCM and 32 KB DTCM single-cycle SRAM (tightly-coupled, cache-less) and AXI4 master/slave interfaces for integration and CPU configuration. The repo includes a developer toolchain and tests (Bazel, Python support) with examples and a Verilator simulator for building and running inference binaries, making it practical for prototyping and integration into low-power ML hardware stacks.