Renesas and Fixstars Unveil Cutting-Edge Tools to Revolutionize ADAS Software

Renesas, a semiconductor company, and Fixstars, which specializes in accelerating multi-core CPU, GPU, and FPGA technologies, are working together to create tools that optimize and simulate software for autonomous driving (AD) and advanced driver-assistance systems (ADAS).

These new tools are designed specifically for Renesas’ R-Car system-on-chip (SoC) devices. They enable developers to quickly build network models that accurately recognize objects right from the start of software development. This reduces the need for revisions later and helps speed up the overall development process.

Hirofumi Kawaguchi, a vice president at Renesas, mentioned, “We’re committed to creating integrated development environments that allow customers to prioritize software development. By supporting deep-learning models tailored for R-Car, we help our clients accelerate their AD and ADAS projects, cut down on time-to-market, and lower development costs.”

Satoshi Miki, CEO of Fixstars, added that their jointly developed cloud-based evaluation environment, Genesis for R-Car, is already helping engineers assess and choose devices early in the development cycle. He emphasized their ongoing efforts to create new technologies that speed up machine-learning operations, which are crucial for keeping automotive software updated and efficient.

AD and ADAS applications rely heavily on deep learning to accurately recognize objects. This requires substantial data calculations and memory. Since automotive applications run on SoCs with limited processing power and memory, the software needs to be extremely optimized. Additionally, the software has to go through multiple rounds of evaluation and updates to ensure peak performance and accuracy.

To address these challenges, Renesas and Fixstars have introduced several tools:

1. R-Car NAS Neural Architecture Search Tool: This tool generates optimized deep-learning network models specifically for the R-Car devices. It leverages the CNN accelerator, DSP, and memory efficiently, allowing the creation of lightweight models that deliver accurate and fast object recognition without requiring deep expertise in R-Car architecture.

2. R-Car DNN Compiler: This compiler converts optimized network models into executable programs that maximize the R-Car’s performance. It ensures these programs run quickly on the CNN IP and optimizes memory to enhance the speed and efficiency of the limited-capacity SRAM.

3. R-Car DNN Simulator: This allows for rapid simulation of compiled programs on a PC, mirroring the operations on the actual R-Car chip. Developers can quickly verify the program’s functionality and tweak the models if the recognition accuracy is affected, further reducing development time.

Renesas and Fixstars plan to keep enhancing their tools to maintain and improve software accuracy and performance through continuous updates. Their first set of tools, now available, is designed for the R-Car V4H SoC, which offers impressive deep-learning performance of up to 34Tops while being energy efficient.