A few days ago, Google announced in a blog post the launch of Coral NPU , an open-source platform that combines a specialized hardware accelerator for AI with a complete set of software tools.
Coral NPU emerges as a pioneering architecture, conceived in close collaboration between Google Research and Google DeepMind. This joint effort has resulted in a design that places Artificial Intelligence at the center, marking a milestone in the development of the next generation of edge AI systems.
Coral NPU: An AI-First Architecture
The fundamental principle behind the design of the Coral NPU is the ability to keep AI running uninterrupted with minimal power consumption. Therefore, the Coral NPU is designed to run AI applications continuously without compromising the battery life of the devices. Its base version achieves a performance of 512 billion operations per second (GOPS) with a mere few milliwatts of power consumption, a remarkably low figure compared to other accelerators on the market.
This architecture has been meticulously optimized to enable all-day AI solutions in wearables and other autonomy-sensitive devices.
Furthermore, its architecture allows manufacturers to adapt and modify the design according to the needs of their own systems-on-a-chip (SoCs). Synaptics will be the first company to integrate this technology into its new Astra SL2610 series of processors, designed for IoT devices, incorporating the Coral-based Torq NPU subsystem.
The flexibility and efficiency of the Coral NPU make it an ideal tool for image, audio, and recognition processing tasks . Among its most notable applications are visual reconstruction, facial and object recognition, real-time translation and transcription, and gesture or voice control. It can also detect user activity (such as walking, running, or sleeping) and adapt to the surrounding physical environment.
With these capabilities, Coral NPU promises to enable a new generation of consumer smart devices, such as watches, headsets, and augmented reality glasses, capable of running advanced AI directly on the device without relying on the cloud.
Technical power driven by RISC-V
The heart of the Coral NPU is based on the 32-bit RV32IMF_Zve32x RISC-V architecture, accompanied by an AXI4 bus and a four-stage pipeline that enables efficient instruction execution. Its design combines in-order dispatch and out-of-order completion, with a scalar and vector approach to balance performance and efficiency.
The system offers SIMD operations for parallel processing of 128-bit vector data, along with 8 KB of instruction memory and 32 KB of data memory. This configuration is complemented by three main components:
- A RISC-V scalar core, programmable in C, responsible for data flow management and ultra-low power consumption.
- A SIMD vector coprocessor, compatible with RVV v1.0 extensions, designed to handle large volumes of information in parallel.
- A matrix coprocessor, optimized to accelerate essential multiply-add (MAC) operations in neural networks.
Open and compatible development ecosystem
Google has accompanied the launch of Coral NPU with a complete development environment that includes AI model compilers (IREE and TFLM), a C compiler, and a simulator for developers.
The NPU is compatible with the most popular AI frameworks , such as TensorFlow, JAX, and PyTorch, and allows models to be compiled into a universal intermediate representation, which is ultimately translated into RISC-V instructions by the LLVM backend.
With this initiative, Google is not only driving the adoption of open hardware for AI, but also laying the foundation for a collaborative ecosystem where manufacturers, developers, and the open source community can innovate on a common platform.
If you are interested in learning more about it , you should know that this project is licensed under Apache 2.0, and you can consult the documentation and tools that Google offers for developers and designers.