Coral NPU: A full-stack platform for Edge AI (developers.googleblog.com)

Google’s Coral team (with Google Research and DeepMind) announced Coral NPU, an open, full‑stack neural processing unit architecture and toolchain designed to bring always‑on, private AI to battery‑constrained edge devices. Coral NPU flips traditional chip design by prioritizing a matrix‑centric ML engine over scalar CPUs, providing a RISC‑V ISA‑compliant, C‑programmable reference NPU that targets ~512 GOPS while drawing “a few milliwatts.” The platform includes an end‑to‑end software stack (TFLM, MLIR/StableHLO, IREE-based compiler plugins, custom kernels, simulator) to streamline progressive lowering from TensorFlow/JAX/PyTorch models to compact edge binaries, and it supports dynamic operators and transformer workloads for small on‑device LLMs. This matters because it tackles the core edge trade‑offs: power efficiency, programmability, and ecosystem fragmentation. By offering an open, extensible hardware IP and a unified developer experience, Coral NPU aims to simplify building ambient sensing, wearables, hearables, AR glasses and IoT devices that can run privacy‑sensitive models locally. Hardware security features (CHERI support) and a production partner implementation—Synaptics’ Astra SL2610 with the Torq NPU—underscore real‑world readiness. For ML engineers and SoC designers, Coral NPU provides a validated path to deploy encoder/transformer models on ultra‑low‑power silicon with standard compilers and a consistent toolchain.