google-coral/coralnpu

A machine learning accelerator core designed for energy-efficient AI at the edge.

What it solves

Coral NPU is an Open Source IP for a hardware accelerator designed to enable machine learning inferencing on ultra-low-power System-on-Chips (SoCs). It specifically targets wearable devices like smart watches, AR glasses, and hearables where power efficiency is critical.

How it works

It is a neural processing unit (NPU) based on the 32-bit RISC-V Instruction Set Architecture (ISA). The architecture consists of three integrated processor components—matrix, vector (SIMD), and scalar—that work together to process ML workloads. It utilizes AXI4 bus interfaces to interact with external memory and allows external CPUs to configure the unit.

Who it’s for

Hardware engineers and SoC designers who want to integrate a dedicated AI accelerator into low-power wearable hardware.

Highlights

  • RISC-V Based: Uses the RV32IMF_Zve32x instruction set.
  • Memory Efficiency: Features 8 KB ITCM and 32 KB DTCM single-cycle-latency SRAM for faster, more efficient access than standard cache.
  • Processing Power: Includes a four-stage processor with four-way scalar and two-way vector dispatch.
  • SIMD Support: Features 128-bit SIMD (with a 256-bit pipeline for the future).

Related

  • Project
    ztachip/ztachip

    An open-source RISC-V AI accelerator for FPGAs and ASICs that optimizes memory bandwidth to efficiently run computer vision and LLM/VLM inference on edge devices.

  • Project
    THU-DSP-LAB/ventus-gpgpu

    Ventus is an open-source GPGPU processor based on the RISC-V ISA and its vector extension, designed for high-performance computing and FPGA deployment.

  • Project
    ucb-bar/chipyard

    An open source framework for the agile development of Chisel-based RISC-V systems-on-chip, integrating processor cores, accelerators, and VLSI tooling.

  • Project
    google/XNNPACK

    XNNPACK is a highly optimized library of low-level performance primitives used to accelerate neural network inference across ARM, x86, WebAssembly, and RISC-V platforms.

  • Project
    ruvnet/ruv-FANN

    A neural intelligence framework featuring a Rust-based neural network engine, advanced forecasting models, and a swarm orchestration system for CPU-native, ephemeral AI.