Google Research completes 32-bit Coral NPU, releases v1.0 compiler and Coralboards
Google Research announced the completed 32-bit Coral NPU architecture, including a new Matrix Execution Unit it says is configurable up to 1 TOPS and targets 10mW average power for inference, plus Coral Compiler v1.0, an open-source IREE/MLIR-based stack that takes JAX, PyTorch and TensorFlow models to the Coral NPU target. The full 32-bit RTL, including the Milestone 4 matrix unit, is available on GitHub, and Google also unveiled second-generation Synaptics Astra SR200 microcontrollers rated at 400 GOPS, while Synaptics Coralboard SL (1 TOPS) and Coralboard SR kits are sold through DigiKey, which TechNews reported as exclusive global distributor. The performance and power figures remain vendor claims and were not independently benchmarked in the reports.
Google Research argues that an open hardware and software standard is the only way to unify an edge landscape split across hundreds of proprietary chip architectures and incompatible toolchains, so teams can optimize a model once and deploy it across SoC vendors and power envelopes.
Unifying Edge AI: How Coral NPU’s open standard brings portable, energy-viable intelligence from model to silicon - Google Developers Blog
Unifying Edge AI: How Coral NPU’s open standard brings portable, energy-viable intelligence from model to silicon
OCT. 20, 2026
Billy Rutledge Eng Director Google Research
#Google Research#Coral NPU#Synaptics#Synaptics Astra SR200#Coralboard#Grinn#DigiKey#IREE#MLIR