AI Accelerator: DRP-AI

Artificial Intelligence (AI) used in deep neural networks is already providing new value for the IT segment. Although many people expect to implement embedded applications with AI, AI processing requires many calculations, making it difficult to adapt to embedded devices using traditional solutions with CPU or GPU, due to insufficient performance or large power consumption demands. Also, AI is constantly evolving, and new algorithms are developed from time to time.

In the midst of the rapid evolution of AI, Renesas developed the AI accelerator (DRP-AI) and the software (DRP-AI translator) that delivers both high performance and low power consumption, and have the ability to respond to evolution. Combining the DRP-AI and the DRP-AI translator makes AI inference possible with high power efficiency, which the current AI technology is unable to support.

The AI model can be extended with the continuous update of the DRP-AI translator.

• Jump to: Features
• Tool
• Software
• Products

Image

Features of DRP-AI

DRP-AI consists of AI-MAC (multiply-accumulate processor) and DRP (reconfigurable processor). AI processing can be executed at high speed by assigning AI-MAC for operations on the convolution layer and fully connected layer, and DRP for other complex processing such as preprocessing and pooling layer.

Image

For more detailed technical information on DRP-AI, please refer to the following white paper.
White Paper: Embedded AI-Accelerator DRP-AI (PDF | English, 日本語)

While most AI accelerators specialize only in AI inference and rely on the CPU for pre- and post-processing, DRP-AI integrates pre- and post-processing and AI inference into a single DRP-AI hardware to achieve superior AI processing performance.

Image

Tool: DRP-AI Translator / DRP-AI TVM^※1

DRP-AI Translator and DRP-AI TVM are tools that are available to convert trained AI models into a format that can run on DRP-AI. This section describes the features of these two tools.
The DRP-AI Translator is a tool that is tuned to maximize DRP-AI performance. DRP-AI achieves high-speed performance, low power consumption and reduced CPU load by enabling DRP-AI to perform all the operations of an AI model.
DRP-AI TVM applies the DRP-AI accelerator to the proven ML compiler framework Apache TVM^※2」. This enables support for multiple AI frameworks (ONNX, PyTorch, TensorFlow, etc.). In addition, it enables operation in conjunction with the CPU, allowing more AI models to be run.

These two tools can be selected according to the customer's product application.

Image

The table below lists the AI model formats and products (MPUs) supported by each tool.

Image

The table below shows the download site for each tool, a brief description of the tool and information on deliverables.

Software: DRP-AI Support Package

The DRP-AI Support Package provides the driver and guide needed to operate DRP-AI. Download the DRP-AI now to experience the seamless AI development from open software to device implementation.

Image

*1 DRP-AI TVM is powered by EdgeCortix MERATM Compiler Framework
2 For more information on Apache TVM, please refer to https://tvm.apache.org