technology. With breakthrough innovations in devices and widespread use of IT infrastructure, the use of data and Machine Learning are becoming more and more familiar to us. We were among the first to predict such a future. I've been running a machine learning based business since 2012. Since we are now working diligently with our colleagues with diverse skills to meet the needs of our customers, we have sought out the following two approaches; "Development of good machine learning models" and "Development of high-speed, highly efficient hardware IP". By approaching from both software and hardware perspectives, we will make the impossible possible. That future is within our reach. We believe that we can enrich people's lives by developing the key technologies of the future and make them available to the world. CEO Soichi Matsuda 5
raised 2019.10 2020.4 Keynote speaker at COOL Chips23 Ultra low power AI inference accelerator IP official launch「Efficiera®」for commerce 10 7 2012.12 Established LeapMind Corporation (formerly AddQuality Corporation) DeLTA-Lite released, "DeLTA-Project" announced 2018.4 Launch of the Efficiera FPGA Partner Program Blueoil, a software stack for quantization neural networks, released as open source software 12
that are unique to the practical application of machine learning. From LeapMind's Inception to the Present Practical applications of machine learning technology Joint Development Operation Extremely low bit quantization Commercial IP OSS 2012創業 2018 2019 2020 Embedded DL development web service
As a Series B Oct.2017 As a Series A Aug.2016 As a Series C Oct.2019 ITOCHU Technology Ventures, Inc. Visionnaire Ventures Fund Archetype Ventures Intel Capital GMO VenturePartners, Inc. NTT DATA Corporation Innovative Venture Fund Investment Limited Partnership ITOCHU Technology Ventures, Inc. Visionnaire Ventures Fund Archetype Ventures Aioi Nissay Dowa Insurance Co., Ltd. SBI Investment Co., Ltd. Toyota Motor Corporation MITSUI & CO., LTD. 3.5 billion JPY LeapMind has raised a total of approximately 5 billion yen in funding to date. About 1.15 billion JPY About 340 million JPY About
a matching service for engineers and clients based on engineering skill visualization, which was expanded to Singapore, and got his business acquired. He founded LeapMind to create a platform of "compact and simple" deep learning technologies that is easily accessible to anyone, thereby contributing to our society and advancing the world. CTO Hiroyuki Tokunaga He completed his master's program at the University of Tokyo Graduate School of Information Science and Engineering in 2007. Prior to becoming a Director and the CTO of LeapMind in 2018, he worked for Yahoo Japan Corporation, Preferred Infrastructure, Inc., and Smart News, Inc. Chief Research Officer & Chief Scientist Atsunori Kanemura, Ph.D. He received the Ph.D. in Informatics from Kyoto University, Japan. He has held positions at research institutions both in Japan and overseas, and his publications include more than 50 papers, 100 presentations, and a tutorial at AAAI, a flagship academic conference on artificial intelligence. Dr. Kanemura joined LeapMind as an Executive Officer and the Chief Research Officer in 2018 to show the future of intelligent machines embedded in various places of our society. VP of Business Katsutoshi Yamazaki Born in 1970. He holds a master's degree from Keio University and has held business management positions in semiconductor and IP manufacturing in Japan, the U.S. and Europe. He joined LeapMind in 2020 and is responsible for the Efficiera business.
collaborating with the number of projects in the field of image recognition of "deep learning" which is the element technology of AI.In particularly, we are focusing on research and development of edge deep learning, which allows deep learning to run on edge devices. AI Machine Learning Neural Network Deep Learning Edge Deep Learning
is expected that the number of IoT devices steadily increases and the amount of data they handle explodes in the next few years, thus the market for edge deep learning is also expected to expand Global IoT Device Number and Forecast （Billion units）
with GPU Embedding FPGA・ASIC Constant internet connection is not required Cheap device unit price Fast response Low power consumption High device unit price High power consumption Requires the internet Slow response High usage fee
The main target market is area where power-saving and edge solutions are required such as automobiles and security, in short, where demand for edge deep learning is increasing Developing business by targeting low-power edge devices Healthcare Automotive Printer Industrial Agricultural / Construction machinery Smartphone Gaming
not found in cloud or GPU-based inference processing. It is expected to be used in situations where real-time responses are required even in the network environment is not stable or personal information is handled. Independent from bandwidth, latency and reliability of Internet connection Reduces security risks as no external transmission is required Reduces costs of uploading data and using cloud 18
accelerator IP Efficiera Knowledge gained from many projects & support system We are developing our business with our original weight reduction technology for deep learning models, the dedicated circuit design and leveraging the knowledge gained from collaboration with more than 150 companies. Quantization technology in deep learning
supports for challenges in each flow of AI development. Joint planning of Al projects Efficiera FPGA Partner Program Efficiera accelerator IP everaging extremely low bit quantization Knowledge gained through numerous projects Choosing the right Al solution Practical use, mass production
several challenges and barriers to practical application. limited computing resources Trade-off between computation and speed Limited processing power, electric power, and other computational resources available on the device With limited computational resources, there is a trade-off between the amount of computation and the processing speed of inference
doubling technique can recover accuracy. 23 One of the methods of weight reduction for deep learning models, "quantization" that has been pushed to the limit. It leads to solutions for edge deep learning issues such as computational complexity. Solution : Extremely Low Bit Quantization
The reduction in memory usage and computation by reducing the weight of the model leads to power savings and a smaller area of computational circuitry, which is key to achieving fast deep learning inference processing on constrained edge devices. Reduces memory usage Reduces amount of computation Power saving and smaller area of computational circuitry 25
that can be implemented on an FPGA device or ASIC/ASSP device, and is specialized for CNN inference operations. Efficiera addresses various technical challenges including power consumption, cost, and heat dissipation, enabling the rapid introduction of on-device edge AI products to the market. 29
of edge AI era, we have focused our hardware and software research efforts into model weight reduction, dedicated circuit design, and building the knowledge needed to deliver full-package solutions. Energy Efficient Features of Efficiera High Performance Small Footprint Scalable By minimizing the volume of data transmitted and the number of bits, the power required for convolution operations is reduced. By reducing the arithmetic logic complexity, the number of operation cycles is reduced and the arithmetic capacity per area/clock rate is improved. By minimizing the number of operated bits, the circuit area and SRAM size per arithmetic logic unit are minimized. Since the computing performance can be fine-tuned by adjusting the circuit configuration, it is possible to optimize the configuration and maximize the performance of Efficiera according to the task being performed.
BoM cost of products with AI functions by integrating Efficiera on the same SoC FPGA device as the CPU and image input circuit Contributing to the reduction of device development costs by achieving practical arithmetic capacity without using advanced semiconductor manufacturing processes such as 5nm and 7nm The most suitable for adding AI capabilities to existing image processing FPGA designs It is able to build AI solutions by training custom data sets based on Efficiera-optimized pre-trained models Since RTL can be implemented using only standard cell libraries and memory, it can be used for many device designs. 31 Achieved 27.7 TOP/W of computing power in 12nm process *Measurement in development prototype
• Extremely low bit quantization AI inference accelerator IP • Optimized for FPGA implementation, also covers ASIC/ASSP • Deep Learning models optimized for Efficiera • Optimized for FPGA performance range • Trained for typical use cases • We also provide tools that enable customers to perform "fine tuning*". ※ A method to build a new model by reusing a part of an existing model
and completed a joint performance evaluation for the construction of an edge AI system. Joint performance evaluation confirms Efficiera can deliver practical performance EIZO's visibility-enhancing system Board DDR SDRAM SoC FPGA DDR Ctlr CPU On-chip Memory Peripherals Efficiera® By taking advantage of the small footprint of Efficiera, one of its many features, it is possible to utilize the free space of the FPGA that is already installed in the DuraVision EVS1VX to achieve deep learning functions, without adding to or changing the hardware. It contributes to overall system cost reduction by achieving a single chip with FPGA 36
in the free space of an existing FPGA without adding or changing hardware, and yet it can deliver practical performance. By combining conventional functions and deep learning on a single chip in FPGA, we were able to keep system costs low. We had frequent meetings to quickly share problems and progress, and the technical support was fast and thorough. What our customers have to say about Efficiera 37 Efficiera was able to achieve practical performance. It led to cost savings. The support was generous and quick.
FPS Crowd counting 1–2 TOP/s 4–8 TOP/s 8–12 TOP/s 60 FPS Noise reduction 40–50 TOP/s As Efficiera can tune its computing efficiency by selecting the circuit configuration, it can cover not only image recognition tasks, such as object detection, but also the real-time processing of resolution improvement tasks that require a much higher performance range, such as super resolution. Efficiera performance scalability Hazard proximity Detection Watching over people in a nursing home Counting thousands of people in a flash Higher resolution for video footage
Automatic control support for drones Automatic detection of scratches and cracks Danger detection by surveillance camera Foreign object detection Use cases Experience in joint development with partners in a wide range of industries
was launched with the goal of co-creation of on-device AI products and solutions that solve customer issues. By participating in this program, partner will be able to combine your company's products and services with Efficiera in order to develop and provide package services and systems that meet your customer's needs. Realizing edge AI Achieving practical application and mass production of edge Al devices for our customers • Providing low-cost FPGA package solutions • Providing co-created solutions that incorporate IP cores from other companies • Joint promotion FPGA solution provider Partners
researchers Tech venture • Members come from 12 countries • English and Japanese is used interchangeably, and meetings are often in English • Female employees play active roles at every layer and the positions of the organization including managers, researchers, and engineers Diversity • Flex system and flexible work style • Close relationship between board members and employees Freedom 49
experts in their respective fields. While each team has a different development flow, a common thread throughout the company is the presence of the Design Doc. We also host regular events for engineers, such as Office Hour and Engineer MeetUp. Office Hour Employee-sponsored study sessions where they can learn about deep learning and other various topics with in-house experts Engineer MeetUp Event featuring talk sessions from CTO/VPoE, LT from engineers, Q&A and more Design Doc A document that describes what, why and how to make each project
take a break from our normal duties for long periods of time. These events allow you to hack a variety of things using your knowledge and the company's resources, whether it's developing something you love, trying something you've always wanted to do, or teaming up with different people to work on something different. The HackDays blog is also posted: https://leapmind.io/blog/2019/09/03/hackdays-2019-3q/
website. Please check it. https://leapmind.io/careers/ Product Owner Interview General Manager Interview General Manager, Efficiera Division Katsutoshi Yamazaki Efficiera Product Owner, Efficiera Division Takuya Wakisaka I want to contribute changing people's lifestyle with Efficiera I want to contribute making society a place to enjoy diversity with Efficiera
website. Please check it. https://leapmind.io/careers/ A good culture of encouraging commitment to following through a whole project. Researcher, Noise Reduction Team, Model Development Group, Efficiera Division Joel Nicholls Got inspired by courage of the team of LM to take on great visions. Engineer, Business Development Team, Commercialization Group, Efficiera Division Lily Tiong
new system depending on phases! 1 on 1 We conduct individual meetings to encourage open communication and to support employee growth 58 Benefits and System At LeapMind, we are working on creating a system to support employees’ various work-styles Work Style Flexible work-styles such as flex system, remote work, and "refresh" leave Free Drinks Mineral water and coffee on the house Development Environment Support to improve the development environment, such as the supply of 4K monitors, laptops and high-back chairs Education Support Support for necessary expenses to improve skills, such as for taking Coursera courses and purchasing books Office Event Social gatherings and barbecues to share information and invigorate communication among all employees
spread of COVID-19 infections in Japan 59 Countermeasures for COVID-19 Work From Home Recommendations Over 80% of our members are working from home. We use Google meet and Slack for job communication and meetings Allowance for Work From Home In addition to the monthly work from home allowance, there is also a ”Remote Work Device Support” subsidy for equipment used in remote work Introduction of a Full-Flex System Full-flexible system with no core time, allowing you to work at your own pace Hosting In-House Events Remotely Organize online events, such as onboarding orientation, company wide meetings, HackDays and DL Office Hour (As of July 2021, subject to change depending on the situation of COVID-19) Interviews are basically conducted online only. Also, we are unable to project new recruiting from abroad for the time being, however, we have relocation support (Welcome Japan Package) to welcome new members from overseas. We would be very much appreciated if you could consider us again when we resume overseas recruitment!
Japan Package to improve the employee experience to support those who are coming from abroad to join LeapMind Support in obtaining a work visa in Japan Partial expense covered for those moving from abroad Online interviews for those who live far away English support channel for personal inquiries Internal materials are written in both English and Japanese Full expenses covered for Japanese Language Proficiency Test JLPT (N1, 2) Aあ
regardless of age, gender, nationality, or race. We, at LeapMind, want to work with someone who wants to make the most of their abilities and skills and also who can respect diversity. Our backgrounds aside, we all share a common goal. To create innovative devices with machine learning and make them available everywhere We are doing the most exciting things right now at LeapMind. Would you like to join us in achieving LeapMind’s vision?