放眼世界，人工智能已成为引领未来产业变革的战略性力量，其中深度学习作为 CONTENTS 人工智能中炙手可热的领域之一，正在帮助计算机以人类智能相似的方式理解世 界。如何快速了解智能的实质，并生产出各种新的具备智能、替代人类工作的机 目录 器及应用，是目前全人类渴望达成的梦想。为了让这些智能载体尽快走入人们的 生活，需要依赖于底层人工智能处理器硬件解决方案，为不断更新的深度学习算 Discovering the Philosophy behind 法技术提供基础性的高性能计算支持和稳定的驱动力。 Deep Learning Computing 深鉴科技在做的，正是致力于为全球人工智能应用提供更高效、更便捷、更经济 的深度学习平台解决方案，并使之可灵活扩展于服务器端与嵌入式端。 目前，深鉴着眼于智慧城市、数据中心、辅助驾驶等领域，帮助用户为多种智能 应用场景打造稳定高效的解决方案。 深 鉴 助力深度学习， 01 02 03 赋予万物智能！ 核心技术优势 产品及应用 团队组成 It is our vision to be able to power all devices with deep TECHNOLOGIES PRODUCTS TEAM learning intelligence. / P03 / P05 / P11 Artificial Intelligence (AI) has become the key in leading the future of world. As one of the most significant breakthroughs of AI, deep learning is enabling devices to understand this world with humankind intelligence. How to unveil the real secret of human intelligence and perfectly inte- grate these smart machines into human life? It has become a current worldwide mission. To realize it, innovation in fundamental hardware solutions is desired, which could free AI algorithms and applications with more stable and faster computing. In leading this innovation, DeePhi Tech provides the world of AI with 04 05 06 07 more efficient, convenient and economic inference platforms which are able to power deep learning from both embedded end and server-side. 合作伙伴 投融资关系 团队荣誉 公司历程 In particular, our reliable and efficient full-stack solutions enable clients PARTNERS INVESTORS HONORS DEVELOPMENT to do various AI applications, including but not only limited to the fields of public security surveillance, data centers, and autonomous vehicles. / P13 / P14 / P15 / P17

// 03 // 04 全球领导性的硬件加速方案和深度学习平台提供者 核心技术优势 TECHNOLOGIES 深 鉴 科 技 (DeePhi) 成 立 于 2016 年， 作 为 全 球 领 导 性 的 硬 Founded in 2016, DeePhi Tech with recognized leader- 深鉴已发表的深度神经网络压缩与加速技术，获得了国内外 DeePhi’s world-leading technology of compression and 件加速方案和深度学习平台提供者，目前已具备世界顶尖的 ship in deep learning acceleration, provides end-to-end 顶尖学术会议的认可，在人工智能领域有广泛而深入的影响 acceleration in deep learning neural networks has re- 神 经 网 络 压 缩 编 译 技 术、 体 系 结 构 与 指 令 集 结 构 设 计 方 法 solutions utilizing deep compression technology and deep 力。 自主研发的灵活、可扩展的 DPU，包括面向 CNN 处 ceived wide recognition from World’s top academic con- 以及 FPGA/ 芯片系统设计集成的研发能力，形成了以自主 learning processing unit (DPU) design. Through synergis- 理的亚里士多德架构 (Aristotle)，面向 RNN 处理的笛卡尔 ferences that deeply influence AI industry. Products from 研发的深度压缩技术与深度学习处理 DPU (Deep Learning tical optimization of neural network deep compression, 架构 (Descartes)， 以及面向 DPU 的国内首款深度学习开 DeePhi, the CNN Accelerators with Aristotle Architecture in Processing Unit) 为核心，从算法到软件到硬件的完整解决方 compiling tool-chain, DPU design, FPGA development, 发 SDK—DNNDKTM(Deep Neural Network Development Edge, the RNN/LSTM Accelerators with Descartes Archi- 案，让所有亟需解决高算力、低功耗需求的智能化设备提供商 DeePhi is able to provide more efficient, convenient, and Kit)，使得人工智能应用可以在嵌入式端和云端的不同设备中 tecture in Cloud, and the first deep learning development 能够方便使用。 economical deep learning inference platforms to end 灵活地实现。 SDK in China, DNNDKTM, they work together as a flexible customers. system for AI applications in both edge devices and cloud computing. 算法 软件 硬件 系统 嵌入式端 服务器端 Compression Software Hardware System Edge Cloud

// 05 // 06 01 算法压缩 COMPRESSION 03 深度学习开发 SDK DNNDK 深度神经网络模型的计算和权重参数往往存在大量冗余， Redundancy in computing operations and weight parameters of DNNDK ™ (Deep Neural Network Development Kit) 是深鉴科技面向 AI 异构计 在片上系统实现时会出现计算和存储资源上的瓶颈，同时 deep neural network models will form a bottleneck of computing 算平台 DPU 自主研发的原创深度学习开发 SDK，为 DPU 平台各种深度学习应用 也会产生大量的功耗。深鉴原创的深度压缩技术，通过剪 and memory resources for on-chip system implementation, mean- 开发和部署提供的一套全栈式高效工具链，它涵盖了神经网络推理 (Inference) 阶 枝和量化两种方式，分别从模型的权重连接数目以及操作 while, cause considerable energy consumption. To solve the prob- 段模型压缩、编译优化和高效运行时支持等各种功能需求，可显著加速用户各类 AI 的数据位宽两个维度来进行模型压缩优化。对于常见的神 lem, DeePhi Tech has developed Deep Compression technology. 应用研发和面市进程。 经网络模型，在保证模型准确率不受损失的前提下，可将 The technique consists of pruning and quantization methods, and 模型大小压缩 5-100 倍，运行时间压缩 1.5-10 倍。 can compress models by cutting down the number of model weight DNNDK ™ (Deep Neural Network Development Kit)- DeePhi ™ deep learning connections and reducing the bitwidth of operands. This technique SDK, is designed as an integrated framework, which aims to simplify \& ac- is also compatible with most main-stream neural networks, and can celerate deep learning applications’development \& deployment on DeePhi compress model size to 5x-100x, running time to 1.5-10x without DPUTM (Deep Learning Processing Unit) platform. It makes the computing losing models’accuracy. power of DPU become easily accessible through providing an innovative \& productive solution, which covers the phases of algorithm compression, pro- Dense Neural Network Prune Pruned Neural Network Quantize Compressed sparse gramming, compilation and runtime enablement. (FP32) (FP32) Parameter Neural Network Finetune (INT8) 04 笛卡尔架构 DESCARTES ARCHITECTURE Quantize Pruning Activation 笛卡尔架构专门为循环神经网络 (RNN，包括 LSTM) 设计，可以针对经过深度压 (Less number of param) 缩后的稀疏神经网络执行高效的硬件加速推理计算。相比于 CPU 和 GPU，用户采 Quantization 用笛卡尔架构可以得到更低的延迟和功耗，例如将模型稀疏化到 20% 的情况下， (Less bits per param) 可以取得相比于 GPU (Tesla P4) 2 倍以上的加速，相对峰值性能高达 2.5 TOPS。 笛卡尔架构可应用于多种深度学习应用场景，如自动语音识别、自然语言处理等。 DECENT 工作流程 / DECENT Workflow Descartes Architecture is designed for compressed Recurrent Neural Networks 02 亚里士多德架构 ARISTOTLE 30FPS (RNN) including LSTM, which could provide efficient hardware acceleration for the inference of sparse neural networks after deep compression. Users are 亚里士多德架构针对卷积神经网络 (CNN) 而设计。目前， face detection/recognition able to achieve lower latency and power consumption with the help of Des- 卷积神经网络一般用来处理图像相关的智能问题，而此架 cartes architecture compared to CPU and GPU, for example, more than 2X 构灵活与可扩展的特性使它可被应用于各种不同规格的终 9CH speedup could be achieved compared to GPU (Tesla P4) when the model is 端中，例如 : 前端的人脸识别解决方案，可实现 30fps 实 1080P pruned to around 20%, together with the performance of 2.5TOPS. Descartes 时高清视频上的人脸检测识别，服务器端的视频结构化， architecture could support varieties of deep learning applications such as auto- 可实现 9CH 1080p 视频的实时识别。 real-timevideo analytics matic speech recognition and natural language processing. DeePhi’s Aristotle Architecture is designed for Convo- 05 笛卡尔高效语音识别引擎 DDESE lutional Neural Networks (CNN). While currently used for video and image recognition tasks, the architecture DDESE 是深鉴基于 Xilinx FPGA 自主研发的高效端到端自动语音识别引擎，它针 is flexible and scalable for both servers and portable 对深度神经网络（主要是 LSTM），为用户提供软硬件协同设计的快速、灵活、高 devices. Based on the architecture, DeePhi’s embed- 效的推理计算加速解决方案（包括剪枝、定点、编译和 FPGA 执行推理）。通过剪 ded face detection/recognition module has a great 枝，可以在保证精度的同时得到一个稀疏模型（稠密度为 15%~20%），然后将模 performance in 30fps real-time video, as well as video 型的权重和激活值定点到 16bit，使整个模型能够被压缩超过 10 倍，可以借助稀 analytics solution can realize high-performance sur- 疏存储格式编译并部署在笛卡尔平台上，采用 FPGA 执行高效的推理。 veillance in 9CH real-time 1080p video. DDESE is an efficient end-to-end automatic speech recognition (ASR) engine based on the FPGA of Xilinx, which is designed for Deep Neural Networks (especially for LSTM), with the deep learning acceleration solution of algo- rithm, software and hardware co-design (containing pruning, quantization, compilation and FPGA inference) by DeePhi. After pruning, the model is pruned to a sparse one (15%~20% density) with little loss of accuracy, then the weights and activations are quantized to 16bits so that the whole mod- el is compressed by more than 10X and could be easily compiled by CSC (Compressed Sparse Column) format and deployed on the Descartes plat- form for efficient inference with the help of FPGA.

// 07 // 08 产品及应用 PRODUCTS 人脸检测识别模组 Face Detection/Recognition Module 笛卡尔高效语音识别引擎 DeePhi Descartes Efficient Speech Recognition Engine 深鉴科技提供一站式智能 IPC 解决方案。采用具 DeePhi Tech delivers a turn-key AI solution for smart IPC. DDESE 是深鉴科技基于 Xilinx FPGA 自主研发的 DDESE is an efficient end-to-end automatic speech recogni- 有颠覆性的 Aristotle 架构，可实现主流深度学习 DP-1200-F01 solution is able to realize real-time deep 高效端到端自动语音识别引擎，它针对深度神经网 tion (ASR) engine based on the FPGA of Xilinx, which is de- 卷 积网络 CNN 在多种场景下的实时应用。该方 learning CNN processing for various application scenarios 络（主要是 LSTM），为用户提供软硬件协同设计 signed for Deep Neural Networks (especially for LSTM), with 案可为行业用户快速实现基于深度学习算法的智能 based on the Aristotle architecture. The solution combines 的快速、灵活、高效的推理计算加速解决方案（包 the deep learning acceleration solution of algorithm, software IPC，在前端实现人脸检测与识别，支持黑 / 白名 both deep learning algorithm and hardware acceleration, 括剪枝、定点、编译和 FPGA 执行推理等）。目前 and hardware co-design (containing pruning, quantization, 单布控。 by which the smart equipment like IPC can be easily em- DDESE 已经发布在 AWS Marketplace 和华为云市 compilation and FPGA inference) by DeePhi. DDESE has bedded with face detection, recognition and white/blacklist 场中，支持用户测试对比 CPU/FPGA 的性能以及 been officially launched in AWS Marketplace and HUAWEI functions at the front-end. 体验单句语音识别效果。 cloud, users could run the test scripts for performance com- parison of CPU/FPGA and single sentence recognition. DP-1200-F01 DP-2400-F09 / DP-2400-O12 人脸分析解决方案 Face Analytics Solution 视频结构化解决方案 Video Analytics Solution 深鉴人脸硬件分析板卡采用深鉴自主研发高性能 DeePhi Tech delivers high-performance FPGA based CNN 深鉴视频结构化硬件分析板卡采用深鉴自主研发 DeePhi video analytics card solution (DP-2400-O12) with FPGA 架构 -Aristotle，集成了基于深度学习的人 accelerator and face analytic deep learning algorithms on 高性能 FPGA 架构 -Aristotle，集成了基于深度学 high-performance FPGA based Aristotle architecture, in- 脸 智能算法。通过对人脸的检测、质量判断、跟踪、 the platform (DP-2400-F09). The solution provides face 习的车辆检测、人员检测、车辆型号识别、车牌识 tegrates with various of deep learning algorithms includ- 轨迹、识别及比对，单板卡最多可实现对 9 路视 analytic functions including detection, quality checking, 别、人员属性、质量、跟踪等深度学习算法。 可 ing vehicle/pedestrian detection, model and plate number 频流和 110FPS 照片流的人脸检测及识别。 tracking, recognition and target path tracing on 9 channels 实时提取视频中机动车、非机动车、行人等属性的 recognition and attributes recognition. It is able to achieve of video streams or 110 fps images as capture machine. 结构化信息。 real-time structural information analysis of motorized/ non-motorized vehicles and pedestrians from multichannels video stream.

// 09 // 10 深鉴 Aristotle 架构平台 Autonomous Driving Platform 深度学习开发 SDK DeePhi Deep Learning SDK 深鉴视频结构化及人脸识别硬件分析平 DeePhi delivers ultra efficient embedded solutions for several computer DNNDK ™ (Deep Neural Network Development DNNDK ™ (Deep Neural Network Development Kit) - DeePhi ™ 台采用具有颠覆性的 Aristotle 架构，在 vision and machine learning applications based on a disruptive hardware Kit) 是深鉴科技面向 AI 异构计算平台 DPU（Deep deep learning SDK, is designed as an integrated framework, 各种平台上提供高 性能 / 低功耗 / 高性 architecture, Aristotle. Aristotle is a dedicated processing unit designed by Learning Processing Unit， 深 度 学 习 处 理 器） 自 which aims to simplify \& accelerate DL (Deep Learning) appli- 价比的解决方案。Aristotle 架构支持主 Deephi. It offers class-leading computational power to support on-device 主研发的原创深度学习 SDK，涵盖了神经网络推理 cations’development and deployment on DeePhi DPU ™ (Deep 流深度学习卷积神经网络的实时处理， neuron networks applications such as face recognition, video structural and （Inference）阶段模型压缩、编译优化和高效运行时 Learning Processing Unit) platform. It makes the computing 其高效处理能力已在多平台上获得验证。 behavior analysis. 支持等各种功能需求，是一套 DPU 平台集各种深度 power of DPU become easily accessible through providing an 针对神经网络算法计算量大 / 并发度高 学习应用开发和部署的高效全栈式解决方案。 innovative \& productive solution, covering the phases of com- 同时数据吞吐量大的特点，我们定制化 Aristotle has been proved to offer high computational efficiency on different pression, programming, compilation and runtime enablement. 地设计了高效的计算阵列和灵活的数据 hardware platforms. When it comes to the state-of-art machine learning al- 流控制流，其易于扩展的特性可支持 日 gorithms, Convolutional Neural Networks (CNN) and Deep Neural Networks, 新 月 异 的 算 法， 实 现 了 加 速 比 和 通 用 which are both computation and memory intensive, Aristotle’s specially 性、功耗和性能的精心优化和平衡。该 designed powerful computation array and flexible data/control flow enable 架构还融合了深鉴 科技领先的神经网络 it to handle the demanding workload and complexity of modern machine 压缩、定点算法，并提供了完整的 SDK learning algorithm at ultra low cost. Aristotle achieves a diligent balance be- 和编译工具，用户在感受到 Aristotle 硬 tween acceleration ratio and versatility, power efficiency and performance. 件架构高性能的同时更能收获简洁方便 的 开 发 体 验。Aristotle 在 Zynq7020 This architecture is also supported by Deephi’s recognized technology in 平 台 上 峰 值 可 达 230GOPS， 支 持 machine learning specializing in deep compression, compiler toolchain, AlexNet、VGGNet、GoogLeNet、 and system-level optimization, which dramatically simplified work load for ResNet 等几乎所有主流卷积神经网络， application development. By effectively parallelizing computation on work- 实测 效率最高超过 90%。 load-specific hardware accelerator, Aristotle can support most mainstream CNN networks such as VGG, GoogleNet and ResNet at Zynq7020, achiev- ing peak performance of 230 GOPS and over 90% measured efficiency. 深度学习全栈式开发工具链 / Full stack toolchain for Deep Learning ADAS 及自动驾驶平台 Autonomous Driving Platform 深鉴科技为 ADAS 及自动驾驶系统提供基于深度学习 DeePhi Tech provides the world leading vehicle computing plat- 处理器的车载感知处理平台 DP-2400-O12。该平台 form for smart driving scenarios. Based on high-performance 以 FPGA 为核心计算单元，搭载 Aristotle 深度学习 FPGA architecture, our deep learning accelerator platform, DP- 专用处理器架构，支持多种传感器融合（路摄像头、 2400-O12, supports different vehicle sensors input (Camera, 激光雷达、毫米波雷达与超声波等），以及以太网、 Lidar, Millimeter wave radar, Ultrasonic, etc.), Ethernet, CAN- CAN 总线与其他通信协议；同时，结合深度学习模型 bus and other communication protocols, to achieve vehicle \& 压缩技术，以高能效、低功耗为目标，实现多路车辆 pedestrian detection, traffic signal detection and recognition, 及行人检测、车道线检测、交通信号检测识别、语义 semantic segmentation and other real-time applications of 分割等其他 ADAS 及自动驾驶系统的实时应用。 ADAS/Autonomous driving capabilities. 目标检测 / Object Detection 背景分割 / Segmentection 车道线检测 / Lane Dectection

// 11 // 12 深鉴创始人 DEEPHI FOUNDERS 团队组成 TEAM 左一：单羿 Yi Shan 左二：汪玉 Yu Wang 左三：姚颂 Song Yao 左四：韩松 Song Han 深鉴科技的创始团队由斯坦福大学与清华大学的深度学习硬件 DeePhi was founded by a group of deep learning accel- 加速研究者组成，技术核心团队的博士学历占比高达 50%， eration researchers from Tsinghua University and Stan- 并且吸引了来自百度、360、华为、海康威视、风河、大疆、 ford University. Team consists of renowned researchers Facebook、Intel、Nokia、Marvell 和 MediaTek 等公司的行 and experienced professionals known for their pioneer- 业精英的加入，使得整个团队在具备国际性学术影响力的同时 ing work in the field of deep learning. More than 50% of 又具备资深的工业经验。 team members have PHD degrees and most of them joint from Baidu, 360 Technology, Huawei, Hikvision, FengHe, DJI-Innovations, Facebook, Intel, Nokia, Marvel, MediaTek and other great companies. 姚颂 汪玉 Song Yao Yu Wang 联合创始人 , CEO 联合创始人 战略，商务，运营 学术界持续性研究，资源整合 毕业于清华大学电子系 清华大学电子系长聘副教授 斯坦福大学电子工程系访问学者 国家自然科学基金“优青”获得者 福布斯亚洲 30 Under 30 清华大学教改以来首批终身教授 MIT TR35 China ACM FPGA 技术委员会亚太唯一成员 IEEE/ACM Senior Member Co-Founder, CEO Strategy, Operation, BD Co-Founder Research, Domestic Resource BS from Tsinghua EE Associate Professor of Tsinghua EE Former Visitng Researcher at Stanford EE NSFC Outstanding Youth Award One of first tenured prof. in Tsinghua Forbes 30 Under 30 Asia ACM SIGDA Reconfigure TC Member MIT TR35 China IEEE/ACM Senior Member 韩松 单羿 Song Han Yi Shan 联合创始人 合伙人 , CTO 学术界持续性研究，海外资源整合 工程，技术 MIT EECS Assistant Professor 清华大学电子系博士 斯坦福大学电子系博士 伦敦帝国理工联合培养博士 Deep Compression 提出者 与 DeepMind 同获 ICLR 2016 最佳论文 IBM PhD Fellowship 入选北京市优秀人才计划 Co-Founder 前百度 IDL 异构计算方向创始成员 Research, Overseas Resource Partner, CTO MIT Assistant Professor Engineering, Product PhD from Stanford EE PhD from Tsinghua EE Proposed ‘Deep Compression’ Joint cultivation in Imperial College London ICLR 2016 Best paper Beijing Excellent Talent Project Founding member of HCG of Baidu IDL

// 13 // 14 合作伙伴 PARTNERS 商业上，深鉴科技凭借领先的技术实力和服务方案，与业内顶 In business, we are striving to bring our cutting-edge solu- 尖科技公司建立了良好的战略合作关系，将我们的系列解决方 tions to all industries utilizing artificial intelligence and have 案应用在安防监控、云服务、机器人、辅助驾驶等智能场景中， established deep collaboration with top companies like 并积极在全球范围内扩展业务。目前的合作伙伴包括 华为、 Huawei、Amazon、Samsung、MediaTek、Ant Financial, Amazon、Samsung、MediaTek、蚂蚁金服、搜狗、东方网力、 Sogou, NetPosa, Genvict, State Grid, Wisesoft, Power- 金溢科技、国家电网、宝德和零度智控等。 Leader and Zerotech. 投融资关系 INVESTORS 2018 年 7 月 18 日 A+ 轮 (2017.10) A 轮 (2017.05) 自适应和智能计算的 全球领导企业赛灵思 天使轮 (2016.03) 获得由蚂蚁金服与三 公 司（Xilinx, Inc.， 星 风 投 领 投， 招 商 局 (NASDAQ:XLNX)） 孵化阶段 (2015) 获 得 金 沙 江 创 投、 高 获得来自赛灵思 (Xil- 创投与华创资本跟投 正式宣布已经完成对 榕资本数百万美元天 inx)、 联 发 科 (Medi- 的 4000 万 美 元 A+ 深鉴科技的收购。 深鉴科技于清华大学 使轮融资。 aTek)、清华控股、方 轮融资。 孵化。 和 资 本， 以 及 天 使 轮 Xilinx, Inc. (NASDAQ: 投 资 方 金 沙 江 创 投、 Closed over $40 XLNX) the leader in Incubated in Tsing- Received millions 高榕资本的数千万美 million round A+ fi- adaptive and intelli- hua University of dollards of angel 元 A 轮融资。 nancing led by Ant gent computing, an- funding from GSR Financial Services nounced that it has Ventures and Ban- Raised tens of mil- Group and Sam- acquired DeePhi yan Capital. lions of round A sung, while China Technology Co., Ltd financing from Xil- Merchants Venture (DeePhi Tech). inx, MediaTek Inc, and China Growth Tsinghua Holdings, Capital participating Sigma Square Cap- as fellow investors. ital, GSR Ventures and Banyan Capital.

// 15 With advanced theories and technologies originating from Tsin- 2017 中关村前沿技术企业 中关村前沿科技与产业服务 // 16 ghua University and Stanford University, DeePhi has an interna- 2017 Z-park Advanced Tech- 联盟理事单位 团队荣誉 HONORS tional influence in both academic and business realms. The in- nology Company Award 第十九届中国国际高新技术成果 novative computing architecture and deep compression method ATECH Alliance Member 交易会优秀产品奖证书 依托清华大学与斯坦福大学的先进学术理念与科技 proposed by DeePhi team has been accept in many academic 2018 中国 IC 设计成就奖 19TH CHINA HI-TECH FAIR 创新方法，在公司创立至今短短的两年时间中，深 conferences such as NIPS, ICLR, FPGA, ISCA, particularly, win 2018 China IC Design Award Excellent Product Certificate 鉴团队在 NIPS、 ICLR、FPGA、ISCA 等顶尖国际 the best paper awards in FPGA 2017 and ICLR 2016. Since 学术期刊和会议上发表了一系列论文，并在 ICLR DeePhi has been founded, we achieved more than 60 local and ChinaBang Awards 2018 EETimes“2017Silicon60” 机器之心 2017 全球 30 大 2016 与 FPGA 2017 会议中获得最佳论文奖。 自 international invention patents, especially in the field of sparse 年度 “人工智能创新”奖 最具潜力半导体创新入选 最佳 AI 创业公司 成立以来，深鉴科技申请中外发明专利 60 余项、收 neural networks, in addition to more than 20 professional honors ChinaBang Awards 2018 获了国内外 20 余项专业荣誉和奖项。尤其在稀疏神 and awards. The Research of Model Compression and Hardware AI Innovation EETimes“2017Silicon60”The Most Potential SYNCED MACHINE 2017 Global 经网络领域已布局二十多项美国专利，三十多项中 Acceleration for Sparse Neural Network has received the first Semiconductor Innovation Company Award Top30 Startup AI Company Award 国专利，并凭借稀疏化神经网络压缩与硬件加速研 prize of 2017 CCF Science and Technology Award. 36Kr - 新商业 100 年度 究，获得 2017 中国计算机学会科学技术奖技术发 最佳 A 轮公司 明一等奖。 36Kr-wise Top100 Round-A Financing Company 新浪科技 - 科技风云榜 新智元 2016 中国最具竞争力 DAC 40 岁以下创新者 2017 最佳初创公司 TOP25 AI 创业企业 汪玉 DAC Under-40 Innovators Award SINA TECHNOLOGY 2017 Best AI Era 2016 Top25 Startup Yu Wang Startup Company Award AI Company ICLR 2016 最佳论文奖 FPGA 2017 最佳论文奖 CCF-GAIR 新智造成长榜 投资家网 2017 年度 AI 领域 (International Conference on (International Symposium on 50 强明星企业 最具投资价值企业 TOP10 Field-Programmable Gate) INVESTORS 2017 The Most Invest- Learning Representations) CCF-GAIR Top 50 Star Enterprise ment worthy Enterprise in AI Field ISCA NIPS 中国计算机学会 麻省理工科技评论 35 岁以下 2017 福布斯亚洲 30 (International Symposium (Conference and Workshop on Neural China Computer Federation 科技创新青年 姚颂 Under 30 姚颂入选 on Computer Architecture) Information Processing Systems) MIT Tech Review Innovators 2017 Forbs 30 Under 30 Asia Song Yao Under 35 Song Yao 80 余项发明专利 30 余项专业荣誉和奖项 invention patents professional honors and awards

// 17 // 18 公司历程 DEVELOPMENT 获得 CCF 科学技术奖技术发明一等奖 自适应和智能计算的全球 Recieved the first prize of 2017 CCF science and technology award 领导企业赛灵思公司（Xil- inx, Inc.，(NASDAQ:XL- 深鉴科技成立，同时获得金沙江创投、 深鉴科技“深有所鉴”首次新品发布会，正式发布 6 款深度学习应用产品 NX)） 正 式 宣 布 已 经 完 成 高榕资本数百万美元天使轮融资 DeePhi's first product launching event with six deep learning products 对深鉴科技的收购。 DeePhi was founded with angel funding 深鉴科技深度学习 SDK-DNNDK 上线 ( 发布 ) Xilinx, Inc. (NASDAQ: from GSR Ventures and Banyan Capital DeePhi deep learning SDK-DNNDK released XLNX) the leader in adaptive and intelligent 第一款基于 Aristotle 架构的卷积神经网络 受邀在 Hot Chips 做专场报告 第一款基于 Descartes 架构的 获得 FPGA 最佳论文奖 获得由蚂蚁金服与三星风投领投，招商局创投与华创资本跟投的 4000 万 computing, announced （CNN）处理器完成 Giving First Presentation in 递归神经网络 / 长短时记忆网络 FPGA best paper award 美元 A+ 轮融资 that it has acquired Dee- Phi Technology Co., Ltd Officially releasing Aristotle Architecture Hot Chips (RNN/LSTM) 处理器完成 Closed over $40 million round A+ financing led by Ant Financial (DeePhi Tech). CNN processor Officially releasing Descartes Services Group andSamsung, while China Merchants Venture and 2016.08 2016.03 Architecture RNN/LSTM China Growth Capital participating as fellow investors processor 2017.02 2017.10 2018.07 2017.01 2015 2016.05 2016.12 2017.05 2018.03 深鉴科技于清华大学孵化。 获得 ICLR 获得最佳论文奖 获得 NIPS Workshop 最佳论文提名奖 获得来自赛灵思（Xilinx）、联发科（MediaTek）、 2018.03 深鉴语音识别加速引擎 DDESE 正式上线亚马逊 Incubated in Tsinghua Best Paper Award in ICLR Best Paper in NIPS 2016 Workshop 清华控股、方和资本，以及天使轮投资方金沙江创投、 AWS，开启全球用户云加速服务模式 University 高榕资本的数千万美元 A 轮融资 2018.03 Officially launched an automated speech recognition engine-DDESE，based on the FPGA Raised round A financing from Xilinx, MediaTek Inc, Tsinghua Holdings, Sigma Square Capital, platform at Amazon AWS. GSR Ventures and Banyan Capital