Google publicly releases complete 32-bit Coral NPU design and launches Coral Compiler v1.0

On October 20, 2026, Google Research announced that it has released the complete 32-bit design of the Coral NPU, which it is developing for low-power edge AI, including the matrix execution unit. It also released Coral Compiler v1.0, which handles AI model conversion and optimization. On the same day, U.S. semiconductor maker Synaptics also announced Coralboard SR, a development board adopting the Coral NPU, and began general sales of Coralboard SL.

Unifying Edge AI: How Coral NPU’s open standard brings portable, energy-viable intelligence from model to silicon - Google Developers Blog

The Coral NPU is an open-source hardware design for running AI model inference on wearable and IoT devices. It is based on the 32-bit RISC-V instruction set and can be incorporated into semiconductor makers' own SoCs. Google announced it in October 2025 as a platform including hardware design and development tools.

In edge AI, chips and development tools differ by manufacturer, so there is a burden of repeatedly adjusting and optimizing models each time the chip used changes. The Coral NPU aims to use a common architecture and open-source tools to make it easier to use a model optimized once on chips and devices from different manufacturers.

Also, when processing that the NPU does not support is left to the CPU, latency and power consumption can increase. Google has designed the Coral NPU hardware and compiler together from the start so that the processing performance needed for devices can be obtained with limited power.

Matrix execution unit released; 32-bit design now complete

Google has developed the 32-bit design in four stages: a scalar core, a vector execution unit, a memory subsystem, and a matrix execution unit. In October it released the Matrix Execution Unit, the final stage, and says all four stages are complete.

According to the datasheet, the matrix execution unit performs 256 multiply-accumulate operations per cycle. Google explains that the design reuses weights and intermediate data used in computations within the unit to suppress power consumption associated with data movement to and from memory.

According to the roadmap, the Coral NPU's computing performance is up to 1 TOPS depending on the configuration, and the target average power consumption during inference is 10 mW.

The design data (RTL) describing the circuit operation is published on GitHub.

Coral Compiler v1.0 handles model conversion and optimization

Coral Compiler v1.0 is an open-source tool that converts AI models into a form that can run on the Coral NPU. According to Google, it supports models created with JAX, PyTorch, and TensorFlow [1].

The compiler infrastructure uses IREE and MLIR, and allocates model processing to the host CPU and NPU. It also performs optimizations to improve runtime efficiency, such as splitting matrices so they fit in the NPU's internal memory.

Using a simulator or FPGA reference design, developers can evaluate model execution performance before manufacturing a chip and adjust the hardware configuration.

Coralboard SR announced; SL now on general sale at DigiKey

Coralboard SR is a microcontroller-based development board for prototyping battery-powered devices such as wearables and industrial sensors. It is equipped with a microcontroller from the Astra SR200 series with a 400 GOPS Torq NPU. The Torq NPU incorporates the Coral NPU's vector execution unit. According to the developer page, its release is planned for the future.

Coralboard SL is a development board for Linux-based IoT devices and robots that lets users try AI processing such as image processing. It is equipped with the Astra SL2619 processor, which has a built-in 1 TOPS Torq NPU. This Torq NPU combines the scalar core of the Coral NPU with Synaptics' proprietary compute core.

After being offered as a limited edition at Google I/O, the Coralboard SL went on general sale on October 20 through the electronic components distributor DigiKey. It can be purchased worldwide, and as of October 22, the price on DigiKey's US site was $99.79 per unit.

Synaptics provides an MLIR-based Torq compiler and runtime for both boards, and says models can be optimized in a common development environment and run on the SL and SR respectively.

For information on each board and development materials, refer to the Coralboard product pages.