资源描述
,AI-Optimized Chipsets,Part I: Key DriversMar 2018,Source: When Moores Law Met AI by Azeem on Medium | icons8,Businesses are increasingly adopting AI to create new applications to transform existing operations. These include connected devices, autonomous vehicles, on-device personal interfaces, voice interactions and AR.,Applications,Data,Computing Hardware,Up to 30 billion more IoT devices are coming online by 2020, streaming data that helps build smarter objects, homes, inform consumer lifestyle, enhance security and energy management.,This positive, recursive ADAC loop where new applications generate more data, in turn enhancing algorithmic complexity, driving demand for higher computing,performance.,1,2,AI Algorithms3Most breakthrough approaches in deep learning use significant computing power. A neural net might have dozens of connected layers and billions of parameters, requiring a step-wise increase in level of computing power.,4,Businesses are increasingly adopting AI to create new applications, driving the development of AI-optimized chips,The ADAC (Applications Data Algorithms Computing Hardware) Loop,Software,Hardware,Autonomous Driving,Speech Recognition & NLP,Computer Vision,Sensors,Business Intelligence,AI Platform,Data,Path Planning,AI-Optimized Chipsets,Industrial Applications,Robotics,Computing,These new applications are built on other technology and infrastructure layer solutions,Source: Vertex | AutomotiveIQ | Icon 8 | Taranis | Kryon Systems | Horizon Robotics,Sensing uses advanced computer vision and perception.Visual tasks including lane detection, pedestrian detection, road signs recognition and blind-spot monitoring are handled more effectively with deep learning.To date, deep learning technologyhas primarily been a software play.Existing processors were not originally designed for these new applications.Hence the need to develop AI- optimized hardware.,Path planning: Simple machine learning algorithms are sufficient to handle driving in high resolution mapped cities or along fixed routes. Deep learning is more suitable in complex situations, (e.g. multiple unknown destinations or changing routes).Software,Taranis offers a comprehensive and affordable crop management solution, and the pest and disease prediction algorithms using deep learning to continually improve accuracy.,Kryon Systems delivers innovative, intelligent Robotic Process Automation (RPA) solutions using patented visual and deep learning technologies.,Horizon Robotics is the leader of embedded AI with leading technologies in autonomous driving perception and decision-making, deep learning algorithms and AI processor architecture.,Examples of Vertex Portfolio Companies that employ deep learning in their solutions,Hardware,Application,Technology,Infrastructure,That may reside in the cloud, on edge devices or in a hybrid environment,Source: : Moor Insight & Strategy,Autonomous VehiclesIn an autonomous car, cameras will generate between 2060 MB/s, radar upwards of 10 KB/s, sonar 10100 KB/s, GPS will run at 50 KB/s, and LIDAR will range between 1070 MB/s.Each autonomous vehicle will be generating approximately 8GB/s, 4TB per day.Autonomous vehicles require a reliable solution withan ultra-low latency of 1ms.,AgricultureDescartes Labs uses deep learning to process satellite imagery for agricultural forecasts.It processes over 5TB of new data every day and references a library of 3PB of archival satellite images.By using real time satellite imagery and weather models, Descartes Labs provides highly accurate weekly forecasts of US corn production compared to monthly forecasts provided by the US Department of Agriculture.,Source: NovAtel,Source: Descartes Labs,And all point to significantly higher data generation,Source: : Intel | IEEE Spectrum, | Deep Learning: An Artificial Intelligence Revolution by Ark Invest | Descartes Labs | Reducing 5G Latency Benefits Automotive Safety by Bill McKinley | NovAtel,500-1000 ms,200 ms,100 ms,1 ms,2G GSM | GPRSEDGE | CDMA 1990- 2000,3G UMTSCDMA 20002000-2010,4G LTE LTE-A2010-2020,5G2020,Source: Wi360,50BNumber of IoT devices by 2020,Smart Home,Wearables,Connected Industries,ConnectedCar,Smart City,Smart Energy,The 5G Evolution: Latency for Different Generations of Cellular Networks100KB/s384KB/s-2MB/s150KB/s-450MB/s10 GB/s,Coupled with the growth of IoT and 5G networks, a data deluge of high volume, velocity and variety is expected,IoT and Exponential Growth in DevicesTop IOT Applications,The growth of IoT and 5G networks expected to generate a data deluge of high volume, velocity and variety,Source: Gartner,Volume,Velocity,Variety,Source: : IoT Analytics | Intel | IEEE Spectrum, | Deep Learning: An Artificial Intelligence Revolution by Ark Invest | Wi360 | icon8,Source: World Economic Forum,Source: Andrew Ng, Ark Invest,Unlike other machine learning algorithms, those associated with deep learning scale with increasing training data,Compounding the power of deep learning, the neural nets themselves have become larger and more sophisticated, as measuredby their number of free “parameters”.Parameters are dials used to tune the networks performance. Generally, more parameters allow a network to express more states and capture more data.It endows computers with previously unimaginable capabilities - understanding photos, translating language, predicting crop yields, diagnosing diseases etc. Enabling AI to write software to automate business processes that humans are unable to write.“The process could be verycomplicatedAs a result of this observation, the AI software writes an AI software to automate thatbusiness process. Because we wont be able to do it. Its too complicated.For the next couple of decades, the greatest contribution of A.I. is writing software that humans simply cant write. Solving theunsolvable problems.”,Jensen HuangCEO | NVIDIA,Source: Inside Microsofts FPGA-Based Configurable Cloud by CTO Mark Russinovich | Nvidia | Deep Learning: An Artificial Intelligence Revolution by Ark Invest | Fortune,Deep Learning vs. Other Programming Techniques,Given future process complexities, AI will be needed to automate the programming process by coding dynamically,Source: Ark Invest Management LLC, Yoshua Bengio,Output,Input,Output,Input, Data Trained Program,Input,Output,Hand Crafted Program,1980s Classic ProgrammingSoftware developer codes the solution in software, which then gets executed in a deterministic and obtuse fashion.This works for simple, well-defined problems but breaks down for more complex tasks.,2000s Machine LearningImproves upon classic programming by replacing some stages of the program with stages that can be trained automatically with dataEnabling computers to perform more complex tasks (e.g. image and voice recognition).The software developer focuses less on coding, more on building models which require enormous datasets to recommend a best output.,2010s Deep LearningEntire program is replaced with stages that can be trained with dataPrograms can be far more capable and accurate.Requires less human effort to create.,Source: Deep Learning: An Artificial Intelligence Revolution by Ark Invest | icon8,Source: Morningstar | Vertex | icon8,But existing processors were not originally designed for new AI applications. Hence the need to develop AI-optimized hardware,Looking ahead,This is the end of Part I of a 4-part series of Vertex Perspectives that seeks to understand key factors driving innovation for AI-optimized chipsets, their industry landscape and development trajectory.In Part II, we review the shift in performance focus of computing from general application to neural nets and how this is driving demand for high performance computing. To this end, some startups are adopting alternative, novel approaches and this is expected to pave the way for other AI-optimized chipsets.In Part III, we assess the dominance of tech giants in the cloud, coupled with disruptive startups adopting cloud-first or edge-first approaches to AI-optimized chips. Most industry players are expected to focus on the cloud, with ASIC startups featuring prominently in the cloud and at the edge.Finally in Part IV, we look at other emerging technologies including neuromorphic chips and quantum computing systems, to explore their promise as alternative AI-optimized chipsets.We are most grateful to Emmanuel Timor (General Partner, Vertex Ventures Israel) and Sandeep Bhadra(Partner, Vertex Ventures US) for their insightful comments on this publication.Do let us know if you would like to subscribe to future Vertex Perspectives.,Source: Vertex,THANK YOU,12,
展开阅读全文