This page provides answers to frequently asked questions about the Coral NPU
platform.

## Why does Google Research develop and offer this NPU platform?

The vision for the Coral NPU project is driven by market feedback and new
technological opportunities. Our experience with the legacy
[Coral](https://gweb-coral-full.uc.r.appspot.com/) products highlights a clear
demand for a platform supporting ML frameworks beyond LiteRT (TensorFlow Lite).
At the same time, the rise of Large Language Models (LLMs) presents a
significant opportunity to bring powerful generative AI to edge devices
efficiently.

Project Coral NPU leverages Google Research's foundational work in open ML
compilers (like IREE) and enhanced AI security, integrating these advancements
into a unified, **open-source** platform. We believe an open approach is the
best way to build a robust and innovative edge AI ecosystem, encouraging wide
adoption and industry-wide collaboration.

## Is Coral NPU meant to be the next generation Edge AI product to replace Coral?

It's helpful to think of Coral NPU as an evolution of Google's edge AI strategy,
rather than a direct replacement for the Coral Edge TPU.

- **Coral** is a Google product suite of hardware modules built around our proprietary Edge TPU chip.
- **Coral NPU** is an open-source NPU architecture that allows silicon partners and chip companies to build their own next-generation, energy-efficient AI accelerators.

The lessons learned from the original Coral project have been applied in Coral
NPU, especially the need for broader ML framework support and a more open
approach.

## What's the migration path for current Coral users to use Coral NPU-based hardware?

The migration path for Coral users is straightforward due to model
compatibility. Coral NPU is an open platform that supports multiple ML
frameworks, including TensorFlow; your existing models should run on new Coral
NPU-based hardware either "as-is" or with minimal conversion.

This migration will be possible once our silicon partners have manufactured and
released their new NPUs or SoCs. When that hardware is commercially available,
you can run your existing models on it.

## Are all the tools and designs in Coral NPU open source?

Yes, the Coral NPU platform is fully open source. This includes the NPU
architecture, which is based on the open RISC-V standard, as well as the
complete MLIR-based software toolchain. All code, documentation, and libraries
are publicly available via our website and open-source repositories.

## What operating systems does Coral NPU work with?

The Coral NPU runs bare-metal programs, targeting RV32IMFV (Zve32x). For
developers working on the Coral NPU RTL, or wanting to run simulation, Debian
Trixie is well tested, though other Linux flavors may work. Coral NPU does not
impose any other restrictions on the operating system other cores in a host SoC
may run.

## Can my company use Coral NPU for commercial products and services?

Yes. Coral NPU is designed for broad commercial adoption. Any company or
developer is free to build and sell commercial products or services based on the
Coral NPU architecture and software toolchain.

## Are license fees paid to Google if my product uses Coral NPU?

**No** , there are no license fees payable to Google for using the Project Coral
NPU architecture or software. The platform is released under permissive
open-source license [**Apache
2.0**](https://www.apache.org/licenses/LICENSE-2.0).

However, it is important to distinguish Google's open-source IP from third-party
IP. If an SoC vendor chooses to integrate the Coral NPU with other proprietary,
licensed components (such as a specific CPU core or memory fabric from another
vendor), those third-party components would still be subject to their own
respective licensing fees.

## How does the NPU differ from a TPU, GPU, and DSP?

These are all types of processors that have different goals, architecture and
optimization for different types of mathematical computations

- **TPU (Tensor Processing Unit)**: Google's proprietary ML accelerator for
  the cloud. TPUs are designed to accelerate model training and
  high-performance inference in data centers.

- **GPU (Graphics Processing Unit)**: The industry standard for training
  complex AI models due to its massive parallelism and high memory bandwidth.
  GPUs are also widely used for high-throughput inference in servers and data
  centers.

- **DSP (Digital Signal Processor)**: A processor for performing real-time
  math on signals (for example, audio). It can run simple, low-latency
  inference tasks but is not designed for the heavy computation required for
  training.

- **NPU (Neural Processing Unit)**: The general term for a processor that is
  highly specialized for efficient AI inferencing on edge devices where power
  consumption is a key requirement.

## What are the industry resources that I can leverage to help me integrate Coral NPU?

The primary resource is our growing open-source ecosystem, which includes
hardware partners, chip design services, and the software toolchain. We are
actively collaborating with silicon partners to integrate the Coral NPU
architecture into their commercial chips. For example, our launch partner
Synaptics is releasing its Torq NPU, an SoC solution ideal for next-generation
IoT and wearable devices.

For a complete list of available hardware, design partners, and software
resources, please visit our [Partners page](https://developers.google.com/coral/partners.md.txt).

## Are there business services available that can help me turn my Coral NPU-integrated design into chip design?

Yes. As part of our ecosystem, we are partnering with leading chip design
service companies. For example, Verisilicon, a leader in RISC-V chip design,
provides expert services to help vendors integrate the Coral NPU into their
custom products. For more details, please refer to our [Partners
page](https://developers.google.com/coral/partners).

## Is it possible to design an AI/ML accelerator chip targeting the IoT device market?

**Yes, absolutely**. The Coral NPU architecture is scalable to meet the diverse
needs of the IoT market. While initial implementations target performance in the
hundreds of GOPS (giga operations per second), the architecture allows partners
to implement more powerful designs targeting the multi- TOPS (trillions of
operations per second) range for more demanding use cases.

<br />


*** ** * ** ***

<br />