本文へジャンプ

成果報告書詳細
管理番号20190000000437
タイトル*平成30年度中間年報 「高効率・高速処理を可能とするAIチップ・次世代コンピューティングの技術開発 革新的AIエッジコンピューティング技術の開発 進化型・低消費電力AIエッジLSIの研究開発」
公開日2019/6/22
報告書年度2018 - 2018
委託先名株式会社ソシオネクスト ArchiTek株式会社 株式会社豊田自動織機
プロジェクト番号P16007
部署名IoT推進部
和文要約
英文要約Title: Project for Innovative AI Chips and Next-Generation Computing Technology Development/Development of innovative AI edge computing technologies/Updatable and Low Power AI-Edge LSI Technology Development (FT2018-FY2020) FY2018 Annual Report

Purpose:
AI technology is rapidly progressing, and now the AI systems is changing from the cloud-centric processing to edge distributed processing. This will revolutionize social infrastructure and industries in such as automobiles, surveillance cameras, robots and other embedded devices. In the past,
AI technology was mainly built on GPU processing for applying to various algorisms and applications. However, for embedded edge applications, the GPU solutions has the problems of power consumption and cost issues.
The purpose of this project is to create the technology for an evolvable/updatable, low power and cost effective AI-Edge solution that can catch up to the new algorism and be applied to the various of applications. The solution is based on "Virtual Engine Architecture" technology that provides the flexibility of programmable hardware accelerator for imaging and AI processing.

Progress:
This year, we worked on the technology development below.
(1)ArchiTek Corporation is working on the development of the "Virtual Engine Architecture".
They extended the architecture for AI processing including image processing, and designed its circuit and have started the evaluation.
Also, they designed the tool for describing the algorisms for the Virtual Engine and finished the alpha version.

(2)TOYOTA INDUSTRIES CORPORATION is working on the development of the real time SLAM technology.
They analyzed the SLAM software and identified the heavy load processing functions and designed the SLAM library which offloads them to the Virtual Engine, and finished the alpha version.
Also, they built the test system using FPGA board, and have started the evaluation.

(3)SOCIONEXT Inc. is working on the development of the Quantized DNN technology and
the first sample of the AI-Edge LSI.
They designed the first model of Ternary and Binary DNN, and released the designed data and software for the FPGA board to HOKKAIDO UNIVERSITY. HOKKAIDO UNIVERSITY evaluated the algorism of the model and made the first feedback.
Also, SOCIONEXT Inc. designed the 8bit Fixed DNN and finished the alpha version.
Regarding the first sample of the AI-Edge LSI, they investigated the requirements of the focused edge applications and defined the LSI specification and the architecture. And they finished the system design of the LSI and the alpha version of the RTL.

As described above, we have worked on the development on schedule. And next year, we will develop the first sample of the AI-Edge LSI and start the each evaluation.
ダウンロード成果報告書データベース(ユーザ登録必須)から、ダウンロードしてください。

▲トップに戻る