News Release2020.6.18New Energy and Industrial Technology Development OrganizationSocionext Inc.ArchiTek Co., Ltd.Toyota Industries CorporationAI edge LSI achieves AI recognition and image processing efficiency 10 times and SLAM time 1/20― Realized by hybrid quantized DNN technology and evolutionary virtual engine architecture technology ―NEDO is engaged in the research and development of advanced and low power consumption AI edge LSIs.OneNext, ArchiTek Co., Ltd. and Toyota Industries Corporation are hybrid quantizers that perform AI recognition processing in the same business.DNN technology, evolutionary virtual engine architecture technology (aIPE) for image processing, and real-timeDeveloped SLAM processing technology. Evolutionary low-power AI edge LSI prototype incorporating these technologiesWhen evaluated, AI recognition processing and image processing are more than 10 times more efficient than general-purpose GPUs, respectively.Real-time SLAM self-position estimation processing time has been shortened to 1/20 compared to CPU.In the future, this AI edge LSI will be used in logistics, machine vision, security/monitoring, and in-vehicle sensing systems.Edge computing that meets the requirements of low power consumption, low latency, and low cost by being applied toIt is possible to build a swing system. As a result, promotion of advanced utilization of rapidly increasing data is acceleratedAs an AI edge technology that realizes decentralization of processing at the edge (edge) side of the network, ultra low power consumptionWe can expect the realization of a power society.Figure 1 Prototype evolutionary low power consumption AI edge LSI1. OverviewIn order to promote the advanced utilization of such data as the amount of data explosively increases due to the arrival of the IoT societyIn addition to conventional cloud data processing, low power consumption at the edge side of the networkIt is required to establish "edge computing technology" for advanced information processing. To realize that,Comprehensive technology from hardware to applications that realize distributed processing on the edge side (total solution"Development of ultra-low power consumption edge computing technology" is required.Against this background, the New Energy and Industrial Technology Development Organization (NEDO) and Socio Co., Ltd.Next, ArchiTek Co., Ltd., and Toyota Industries Corporation are the artificial intelligence companies in the NEDO business *1 .(AI) AI processing for the purpose of developing recognition processing, various image processing, and real-time SLAM *2 processing technology.We have promoted research and development themes for technology. And now, a hybrid quantization deep that performs AI recognition processingPage 2Neural network (DNN) technology *3 , evolutionary virtual engine arche that can execute various image processing in parallel at high speedWe have developed texture technology (aIPE) *4 and real-time SLAM processing technology. Progress that incorporates these technologiesAs a result of prototyping a low-power, low power consumption AI edge LSI *5 , it was compared with general-purpose GPU *6 in AI recognition and image processing.Compared to each other, we succeeded in improving the power efficiency by 10 times or more, and in the real-time SLAM self-position estimation processing,The processing time has been shortened by 1/20 compared with the CPU for use.In the future, AI edge LSIs using this technology will be used for logistics, machine vision, security/monitoring, and in-vehicle sensing systems.Edge computing that meets the requirements of low power consumption, low latency, and low cost when applied to systems such asIt is now possible to build a network system, accelerate the promotion of advanced utilization of rapidly increasing data, andIt is expected that an ultra-low power consumption society will be realized as an AI edge technology that realizes decentralization of processing on the side of the edge.2. This achievementThis time, a hybrid developed by NEDO, Socionext Co., Ltd., ArchiTek Co., Ltd. and Toyota Industries CorporationThe dequantized DNN technology, the evolved aIPE, and the real-time SLAM processing technology have the following features.(1) Hybrid quantized DNN technologyReduced the bit *7 and activation *8 parameters required to execute deep learning .It is a technology. The developed hybrid quantized DNN technology uses conventional floating point 32-bit/16-bit and integer 8-bitThe backbone network *9 represents the network structure represented by a quad, and the ternary number is 2 bits, the binary number is 1 bit, and the head.Network *10 is a technology that mixes integer 8 bits and multiple quantization precisions. This reduces the recognition accuracy.We realized low power consumption while suppressing. Furthermore, a quantization library that is compatible with TensorFlow *11 in a learning environment ,We have developed a hybrid quantization engine in the inference environment and a conversion processing technology from the learning environment to the inference environment.Using these technologies, AI recognition processing can be executed at high speed and with low power consumption. Measure the prototype AI edge LSIAs a result, the hybrid quantized DNN technology has more than 10 times more power efficiency in AI recognition processing than general-purpose GPU.Successful rationalization.Fig. 2 Overview of the developed Quantized Deep Neural Network (DNN) technology that performs AI recognition processingPage 3(2) Evolutionary aIPE and real-time SLAM processing technologyTraditional virtual engine architecture technology (aIPE) combines a minimum of hardware components toBuild rhythms and place them densely in the time slots in which they run, for high efficiency, flexibility and versatilityI've been. The newly developed evolutionary aIPE is an architecture for improving optimization and implementing AI extension functions.Image processing is performed with high speed and low power consumption by improving the image processing. The main technologies developed are image processing and real-timeIm SLAM is a technology related to high-speed processing. Due to the architecture that can maximize the SLAM algorithmPerforms SLAM processing using the LiDAR *12 sensor efficiently. In addition to these, ① Hardware arbitration mechanismTechnology to improve and improve the memory control function, (2) Flexible net definition for deep learning is possibleAnd, technology to freely plug in element parts such as convolution circuits, ③ Function from LiDAR to Visual SLAMImprovement, (4) Developed a quantization calculator. Incorporate these technologies into conventional aIPE and endure from logic design to implementationIt becomes a platform that can support algorithm evolution and a wide range of AI edge applications by converting to IPIt was.Figure 3 Overview of virtual engine architecture technology that can execute various image processes in parallelAs a result of measurement with a prototype AI edge LSI, with evolutionary aIPE, 10 times or more compared with general-purpose GPU in image processingWe have succeeded in improving the above power efficiency.We have also developed a real-time SLAM processing library that can operate in parallel on the evolved aIPE of the prototype LSI.For highly accurate self-position estimation processing of high-speed mobile robotsIt was confirmed that the processing time can be shortened to 1/20 of that of the CPU.3. Future plansNEDO and Socionext Co., Ltd., ArchiTek Co., Ltd., and Toyota Industries Corporation will continue to develop advanced aIPE and hybrid quantities.Integrate child DNN technology, along with advancement of real-time SLAM processing technology, computer vision and AI basicsFurther development of middleware library and optimization technology of cloud environment and edge environment,We will build an advanced and low power consumption AI edge LSI that operates with even lower power consumption. This allows industrial inspection,Aiming to establish technology that can execute advanced AI with low power consumption in order to expand its application to driving assistance, drones, etc.Page 4To do.The platform that incorporates the newly developed evolutionary aIPE will be available from ArchiTek Co., Ltd.(Circuit information) will be provided.[Note]*1 NEDO businessBusiness name: AI chip that enables high-efficiency and high-speed processing, next-generation computing technology development/innovative AI edge computingDevelopment of advanced technology/Research and development of advanced and low power consumption AI edge LSIBusiness period: 2018-2020*2 Real-time SLAMSimultaneous Localization and Mapping (simultaneous execution of self-location estimation and environment mapping). Self-localization of moving objects and environmental mappingIt is a generic term for technologies that perform growth at the same time.*3 Quantized deep neural network (DNN) technologyDeep Neural Network (DNN) is a neural network that supports deep learning and has four or more layers deep.Quantized DNN reduces the amount of calculation by lowering the bit of the processing of the DNN calculation algorithm to perform low power consumption calculation.*4 Virtual engine architecture technology (aIPE)Abbreviation for ArchiTek Intelligence® Pixel Engine. The functions required for image processing and AI processing are extracted orthogonally and realized by hardware,A flexible engine that can execute advanced and versatile applications by combining these hardware components.It is a technology originally developed by ArchiTek Co., Ltd. We have rationalized and strengthened hardware parts so that they can be widely used with few resources.In point, it realizes small size, low cost, and low power consumption.*5 Edge LSIA semiconductor chip used in IoT devices located at the end of the network close to users. Use by comparing with cloud or serverIt is a chip with severe restrictions on power, heat generation, and cost.*6 General-purpose GPUGraphics Processing Unit ( arithmetic processor for image processing). A computing device specialized for real-time image processing.*7 ParameterA large number of fixed values used in calculations.*8 ActivationInput value of the intermediate layer in the neural network.*9 Backbone networkA network that extracts feature maps from image input.*10 Head networkThis network receives the feature map from the backbone network and executes post-stage processing.*11 TensorFlowAn open software library for use in various machine learning fields, flowing multi-dimensional data structures (tensors)Is a library for deep learning that can be processed like.*12 LiDARAn acronym for "light detection and ranging," a laser beam is emitted to bounce off an object.It is a technology that measures the time to arrive and measures the distance and direction to an object.Four. Contact information(Contact for inquiries regarding the contents of this news release)NEDO IoT Promotion Department Person in charge: Hirose, Nishiyama TEL: 044-520-5211Page 5Socionext Public Relations Officer: Izumi TEL: 045-568-1006Inquiry Form https://www.socionext.com/jp/contact/ArchiTek Person in charge: CFO Fujinaka E-mail: [email protected] Industries Corporation Public Relations Department Contact: Miyazaki TEL: 0566-27-5157(General inquiries regarding other NEDO projects)NEDO Public Relations Department Contact: Sakamoto, Sato TEL: 044-520-5151 E-mail: [email protected]
BRN Price at posting:
9.3¢ Sentiment: Buy Disclosure: Held
(20min delay)