OKI Electric Industry Co., Ltd.
OKI IDS Co., Ltd.
Mipsology SAS
Tokyo, October 17, 2022 -- OKI Electric Industry Co., Ltd. (President: Takahiro Mori; office: Minato-ku, Tokyo Japan; "OKI"), OKI IDS Co., Ltd. (President: Ienobu Takizawa; office: Takasaki-shi, Gunma Japan; "OIDS") and Mipsology SAS (CEO: Ludovic Larzul; office: France, "Mipsology") have successfully developed the automation technology for the whole processes from AI model streamlining to the implementation in FPGAs based on PCAS (*1), OKI's proprietary solution to compress AI models, and Zebra, Mipsology's FPGA AI acceleration platform. The solution has been confirmed capable of processing AI images at a rate of 60 fps, four times faster than conventional, thus proving how advanced it is for social deployment. Furthermore, the FPGA (*2) design service for Japanese customers by OIDS and Mipsology provides embedded applications equipped with sophisticated AI in a short period of time.
A conceptual diagram of the FPGA deployment
AI development is faced with the requirement for faster (real-time) processing and reduction in power consumption to deploy self-driving, remote-controlled robots, telehealth and video monitoring in edge domains.
As sophisticated AI models (deep-learning models) entail large-scale and complex operations, expert engineers are essential who have specialized knowledge in AI model compression and design of circuits to deploy AI models' arithmetic logics in edge devices with severe restrictions on computation capacity and memory usage. FPGAs can arrange intricate operations in parallel to process a volume of data rapidly while requiring even more complex implementation and expertise than general-purpose CPUs and GPUs, hence longer development time and higher costs.
OKI's developed PCAS compresses AI models for any hardware with the recognition accuracy maintained to the maximum extent. The combination of PCAS with Mipsology's Zebra AI acceleration platform automates the streamlining and deployment of AI models to FPGAs, hence shorter development time and lower costs. A test of AI image processing (YOLOv4 (*3) as the model) confirms the processing rate of 60 fps or more, four times faster than without PCAS compression. The rate meets the requirement to deploy AI applications to society. We can also expect higher speed and lower power consumption.
FPGA inference results of PCAS×Zebra
OIDS initiated the joint solution of PCAS and the FPGA design service for Japanese customers with Mipsology launched in November 2020. OKI's PCAS and Mipsology's Zebra platform will allow customers' AI programs to be seamlessly deployed to FPGAs. We are all aiming to achieve customers' required performance in a short time.
OKI's developed technology to cut the amount of computation of deep learning models.
An integrated circuit that allows buyers or designers to set up their configuration after production. It is a kind of PLDs (programmable logic device).
A deep learning model for object detection in the form of open-source software (whose source code is disclosed free of charge, and use, modification and redistribution are permitted under certain conditions). With its precise detection capability, it has been employed in a wide range of cases.
OKI and Mipsology of France Enter Japanese Market for FPGA-Based Machine Learning (ML) Development Support Services
https://www.oki.com/en/press/2020/11/z20067e.html
Part of the results is the outcome of a commissioned project (JPNP16007) by New Energy and Industrial Technology Development Organization (NEDO).