Contents
Google has released the complete hardware design for its 32-bit Coral NPU, including the matrix execution unit, along with Coral Compiler v1.0. The October 20, 2026 announcement gives chip designers access to a fully open design and developers tools to evaluate models before chips are manufactured.
Key Points#
- The complete 32-bit design includes a scalar core, vector execution unit, memory subsystem, and matrix execution unit.
- According to Google, Coral Compiler v1.0 supports models developed with JAX, PyTorch, and TensorFlow.
- Google's roadmap shows configurations of up to 1 TOPS and targets average power consumption of 10 mW during inference. These are not measured results across all implementations.
- Synaptics announced plans to release Coralboard SR and began general sales of Coralboard SL. As of October 22, SL was listed on DigiKey's US site for US$99.79.
What Google Released#
Coral NPU is an open-source hardware design based on the 32-bit RISC-V instruction set. Semiconductor companies can integrate it into their SoCs to run AI inference on devices such as wearables and industrial sensors.
This release completes all four stages of the design: the scalar core, vector execution unit, memory subsystem, and matrix execution unit. Register-transfer-level design files describing circuit behavior are available on GitHub.
According to the datasheet, the matrix execution unit performs 256 multiply-accumulate operations per cycle. Google says it reuses weights and intermediate data within the unit to reduce the energy spent transferring data to and from memory. This matters because computational throughput alone does not determine whether AI workloads can run on battery-powered devices.
How the Compiler Helps#
Coral Compiler v1.0 converts AI models into a format that runs on Coral NPU. Built on IREE and MLIR, it distributes model operations between the host CPU and NPU and performs optimizations such as dividing matrices into chunks that fit in the NPU's local memory.
Google's goal is to reduce the repeated model conversion and optimization work required when changing chips. A common architecture and open tools could make it easier to move models between implementations. However, developers still need to check operator support, memory requirements, and performance on each target.
The simulator and FPGA reference design let developers evaluate models before chips are manufactured. A practical approach for hardware teams is to compile representative models, check which operations run on the NPU, and verify that the intended configuration meets latency and memory constraints.
Which Board Fits Your Project?#
Synaptics' two Coralboard products address different device requirements. Their Torq NPUs use different parts of Coral NPU, so neither board should be viewed as a direct implementation of the complete design released here.
| Board | Intended Use | Processor and NPU | Availability | |---|---|---|---| | Coralboard SR | Battery-powered wearables and industrial sensors | Astra SR200 series microcontroller with a 400 GOPS Torq NPU incorporating Coral NPU's vector unit | Announced, awaiting release | | Coralboard SL | Linux-based IoT devices, robotics, and image processing | Astra SL2619 processor with a 1 TOPS Torq NPU combining Coral NPU's scalar core with Synaptics' compute core | General sales began October 20 |
As of October 22, 2026, Coralboard SL was listed on DigiKey's US site for US$99.79. Check current pricing, stock, and regional availability before ordering. Synaptics provides an MLIR-based Torq compiler and runtime for both boards and says models can be optimized for SL and SR in a shared development environment.
What Needs Verification#
The roadmap's targets of up to 1 TOPS and average power consumption of 10 mW during inference do not guarantee finished-product performance or battery life. Actual results depend on the hardware configuration, model, memory traffic, and workload. The power target should also not be interpreted as the consumption of an entire development board.
CPU fallback requires particular attention. Moving operations that cannot run on the NPU to the CPU can increase latency and energy consumption. Framework support therefore does not guarantee that an entire model will run efficiently on the accelerator.
Before choosing a chip or board, verify four things: supported operations, peak memory usage, end-to-end latency, and power consumption under the intended inference schedule. Compare these results with your device requirements rather than choosing hardware based on TOPS alone.
Next Steps#
Compile one representative model with Coral Compiler v1.0 and evaluate it in the simulator, recording operations that fall back to the CPU, memory usage, and latency before choosing a hardware configuration.
Source: https://gihyo.jp/article/2026/10/google-coral-npu-32bit-compiler-v1
Contents
