These tools enable the development of highly accurate object recognition network models from the early stages of software development, leveraging the capabilities of the R-Car. This reduces post-development work and thus helps shorten development cycles.
"GENESIS for R-Car, a cloud-based evaluation environment we co-built with Renesas, allows engineers to evaluate and select devices earlier in the development cycle and has already been used by many customers," said Satoshi Miki, CEO of Fixstars. "We will continue to develop new technologies to accelerate machine learning operations (MLOps) that can be used to maintain the latest software versions in automotive applications."
Current AD and ADAS applications use deep learning to achieve highly accurate object recognition. Deep learning inference processing requires massive amounts of data computation and memory capacity. Models and executable programs in automotive applications must be optimized for an automotive SoC, as real-time processing with limited arithmetic units and memory resources can be challenging. Furthermore, the process from software evaluation to verification must be accelerated, and updates must be applied repeatedly to improve accuracy and performance. Renesas and Fixstars have developed the following tools designed to meet these needs:
1. The R-Car Neural Architecture Search (NAS) tool for generating network models optimized for R-Car.
This tool generates deep learning network models that efficiently utilize the R-Car device's CNN (convolutional neural network) accelerator, DSP, and memory. This allows engineers to develop lightweight network models that achieve high-accuracy object recognition and fast processing times, even without in-depth knowledge or experience with the R-Car architecture.
2. R-Car DNN Compiler for Compiling Network Models for R-Car:
This compiler converts optimized network models into programs that can fully leverage R-Car's performance potential. It transforms network models into programs that can run quickly on the IP CNN and also performs memory optimization to allow the high-speed, limited-capacity SRAM to maximize its performance.
3. R-Car DNN Simulator for Rapid Simulation of Compiled Programs:
This simulator can be used to verify program functionality on a PC, rather than on the R-Car chip itself. With this tool, developers can generate the same operational results that R-Car would produce. If the recognition accuracy of the inference process is affected during the model lightening and program optimization process, engineers can provide immediate feedback to the model development team, thus shortening development cycles.
Renesas and Fixstars will continue developing deep learning software with the joint "Automotive SW Platform Lab" and will build operating environments that maintain and improve recognition accuracy and performance through continuous network model updates.
Availability:
The first toolset available today is designed for the R-Car V4H SoC for AD and ADAS applications, combining powerful deep learning performance of up to 34 tera operations per second (TOPS) with superior energy efficiency.
