Renesas and Fixstars, a specialist in multi-core CPU/GPU/FPGA acceleration technology, are to collaborate in the automotive deep learning field.

In April the two companies plan to establish an Automotive SW Platform Lab tasked with the development of software and operating environments for Renesas automotive devices. The new Lab will support early development and ongoing operation of advanced driver-assistance systems (ADAS) and autonomous driving (AD) systems.

The collaboration will look to develop technologies aimed at software development for deep learning and building operating environments that have the ability to continuously update learned network models to maintain and enhance recognition accuracy and performance.

“Fixstars possesses both advanced software technology for deep learning and optimisation technology that allows more efficient utilisation of hardware,” said Takeshi Kataoka, Senior Vice President, General Manager of the Automotive Solution Business Unit at Renesas. “Our collaboration will enable us to provide strong support for software development optimised for automotive applications and allow our customers to fully leverage the superior performance of Renesas’ automotive devices.”

“After developing a deep learning application, it is not possible to maintain high recognition accuracy and performance without constantly updating it with the latest learning data,” added Satoshi Miki, CEO of Fixstars. “Fixstars plans to focus on these machine learning operations (MLOps) for the automotive field, as we work together with Renesas to develop a deep learning development platform optimised for Renesas devices.”

As part of their collaboration, Renesas and Fixstars are launching GENESIS for R-Car, a cloud-based evaluation environment for R-Car that supports early development of ADAS and AD systems. The new environment facilitates instant initial evaluations when selecting devices. It utilises the GENESIS cloud-based device evaluation environment from Fixstars as its platform.

Evaluation based on actual use cases is essential when selecting devices. Users typically need to obtain an evaluation board and basic software to evaluate devices, and technical expertise is also required to build an evaluation environment. The GENESIS for R-Car cloud-based evaluation environment, by contrast, does not require specialised technical expertise.

GENESIS for R-Car lets engineers confirm the processing execution time in frames per second (fps) and recognition accuracy percentage of R-Car V3H’s CNN accelerators on sample images using generic CNN models, such as ResNet or MobileNet. It also allows engineers to select the device and network they wish to evaluate and perform operations remotely on an actual board.

Engineers can use the GENESIS environment to confirm evaluation results in tasks such as image classification and object detection, with the option to use their own images or video data. This greatly simplifies the initial evaluation to determine whether R-Car V3H is suitable for the customer’s system.

Future plans include the rollout of a service that will allow customers to use their own CNN models for evaluations.