R-Car V4H is an automotive SoC that delivers best-in-class deep learning with ultra-low power consumption for autonomous driving levels 2 to 3. This will introduce the tools to get the most out of this R-Car V4H in computer vision for AD/ADAS applications.

AD/ADAS applications often utilize recognition processing based on deep learning to achieve highly accurate image recognition. The volume of operations performed in this recognition process is increasing every year, depending on the amount of information handled and the high level of accuracy required.

Convolutional Encoder-Decoder

On the other hand, since semiconductor devices are entering the post-Moore era, the enormous amount of calculations that continue to increase every year must be achieved with the same power, execution speed and semiconductor process as before. It is impractical to attempt to achieve this with multi-core general-purpose CPUs and GPUs due to their enormous current consumption. This is where heterogeneous AI devices that have multiple dedicated application-specific hardware accelerators are needed. In addition, you need knowledge of specialized tools to use them.


Software List

Designing and verifying neural networks and maximizing the use of dedicated hardware requires specialized knowledge and dedicated tools, so we are introducing the following three solutions to help you with these tasks.

  • R-Car NAS (Neural Architecture Search) is a tool for automatically designing deep learning models that run efficiently on R-Car
  • R-Car DNN Compiler is a DNN compiler that automatically applies program optimization for R-Car V4H deep learning models
  • R-Car DNN Simulator is a high-speed simulator for deep learning model R-Car programs

Overall Configuration


R-Car NAS (Neural Architecture Search) is a tool to generate optimized network models for R-Car

This tool generates deep learning network models that efficiently utilize the CNN IP, DSP and memory in R-Car. This allows early development of lightweight network models that achieve the recognition accuracy and processing time requirements without requiring an in-depth knowledge and understanding of R-Car.

If you input a config file with specifications based on your own dataset, it will automatically generate the neural network that operates most optimally on the R-Car.

Representative Diagram


R-Car DNN Compiler is a tool to compile network models for R-Car

This compiler converts the optimized network model into a program that is able to take full advantage of the R-Car's performance. This converts the CNN IP into a program that can be executed at high speed and optimizes the memory to maximize the use of high-speed, small-capacity SRAM.

We provide this structure by extending our tools to the open source compiler Apache TVM [1].

Representative Diagram


R-Car DNN Simulator is a tool for high-speed simulation of compiled programs on a PC

This simulator enables high-speed software operation verification on a PC without using the actual R-Car chip. There is a phenomenon where the recognition results will differ slightly between your neural network model and the actual hardware output, but this effect can be confirmed by using a high-speed simulator, even without having the hardware.

Representative Diagram

R-Car NAS (Neural Architecture Search)
Tool for automatically designing deep learning models that run efficiently on R-Car
Model-Based Development Renesas
R-Car DNN Compiler
DNN compiler for automatically applying program optimization for deep learning model for R-Car V4H
Compiler/Assembler Renesas
R-Car DNN Simulator
High-speed simulator for deep learning model programs for R-Car
Simulator Renesas
3 items
Tools to Optimize AI Software for AD/ADAS on R-Car SoC