Era ChengWen Wu 吳誠文 09 29 2018 AI Chips Arms Race Source wwwjmyangcomblog2018327aichipstartupmonitor20180328 Nvidia and Google are currently leading the race ID: 816293
Download The PPT/PDF document "Challenges and Opportunities for Future..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Challenges and Opportunities for Future Semiconductor Products in the IOT Era
Cheng-Wen Wu (吳誠文)09/29/2018
Slide2AI Chips Arms RaceSource: www.jmyang.com/blog/2018/3/27/ai-chip-startup-monitor-2018-03-28
Nvidia and Google are currently leading the raceFollowed by Intel, IBM, Qualcomm, NXP, AMD, Samsung, Xilinx, HiSilicon, MediaTek, Bitmain, …And even Microsoft, Amazon, Apple, Facebook, Baidu, Alibaba, Tesla
…
Dozens of
startups
are jumping in—what a tsunami!
Slide3Level 4 Autonomous Vehicle LeadersGoogle (
Waymo)Level 4 pilot services in Phoenix area, with the best AV technology so farNeeds mass production (partners) of the fleet of cars/trucksPartners: Fiat Chrysler, Jaguar, Walmart, Avis, AutoNation, …GM (Cruise)Acquired self-driving car start-up, Cruise Automation, in 2016To build fully autonomous vehicles in a mass-production assembly plant—for ride-sharing firstNeeds cloud/AI/5G technologiesPartners: Softbank, Lyft, …
Slide4WaymoWaymo have been conducting road tests of their self-driving cars and trucks in California and Arizona (Level 4)
They’re headed to Atlanta, Georgia, one of the biggest logistics hubs in the US [announced in March, 2018]Their self-driving cars will serve as a link to public transit in Phoenix [announced 7/31/2018]Source: Waymo & Uber, 2018Google: in 10.5 years
9M+
Miles of Road Test
25K miles/day
6B+
Miles of Simulation
Uber
:
end of July, 2018
“We will shut down our self-driving truck project (Otto)”
Law suit by Alphabet
soon after Uber acquired Otto
Pulled its
robo
-cars from the roads after fatal accident
Slide5Baidu Create ‒ Intel Inside Baidu Create 2018 in Beijing is looking like Google I/O in Silicon Valley
Intel/Mobileye announced that the Responsibility Sensitive Safety (RSS) model will be designed into Baidu’s open-source (Android like) Project Apollo and commercial Apollo Drive programsApollo platform has signed more than 100 car OEMs and tier onesThe RSS model has two separate systems:AI based on reinforcement learning proposes the AV’s next actionSafety layer based on a formal deterministic system can override an “unsafe” AV decisionBaidu was the 1st to announce the adoption of RSSMay also adopt Mobileye’s Surround Computer Vision Kit as its visual perception solutionBaidu’s PaddlePaddle (DL framework) is optimized for Intel XeonOn the shopping list: Intel’s Xeye (AI camera powered by Intel’s Movidius vision processing unit), Intel’s FPGAs, etc.
Source: EE Times, July 4, 2018
Slide6Alexa, Google Assistant, Siri, etc., have become our digital helpersThey set timers, play music, check weather, read news, schedule appointments, teach
kids how to do homework, control your home appliances, …Soon to come with display and other sensorsSmart Speakers in the USSource: www.pcmag.com/article/357520/the-best-smart-speakers, 2/15/2018
Slide7Source: Tiffany Trader, HPC Wire, 4/11/2018
US machine Summit recaptures supercomputing crown [IEEE Spectrum, 6/25/2018]
Slide8Source:
Political Calculations, 8/31/2018
網路泡沫
金融海嘯
Slide9Source: Thomas Piketty, Emmanuel
Saez, and Gabriel Zucman, Distributional National Accounts: Methods and Estimates for the United States, Sept. 2017
Source: TIME, May 28, 2018
Slide10Our Broken Economy
Source: The New York Times, Aug. 7, 2017
Slide11Global Shortage of Energy Supply
In 2014, cloud data centers consumed about 1.62% of global energyAbout 1.8% in the US (70B KWH) [US Data Center Energy Usage Report, LBNL, 6/2016]In 2017, the number was higher than 3%---can rise to 20%
by 2025
[
data-economy.com,
12/2017]
CNBC
interviewed
David Patterson
[www.cnbc.com, 5/6/2017]
“Four years ago, Google worried that if every Android user had 3 minutes of conversation translated a day using machine learning, they'd have to double their data centers.”
Alphabet spend about $10B each year on
Google
data center equipment
New data centers
or
improved equipment
Google
TPU
outperforms CPU by
15-30
X
30-80
X in energy efficiency
Slide12Speech recognition error dropped from 22% (201
3) to 4.9% (2017)Image recognition error (ImageNet contest) dropped from 26% (2011) to 3.5% (2016)Key success factors:Effective training dataScalable DNN modelEfficient hardwareTensorFlow: Open source ML platformCloud ML: ML on any dataset, of any sizeCloud AutoMLBuilt on Google’s transfer learning and neural architecture search technologies (among others)TPU usage fee: $6.50/hr
Google: DNN
Is Proven Effective
Source:
Google, 2017, 2018
Slide13Competition in AI Cloud Services
Amazon, Google, and Microsoft all want to dominate cloud AI servicesFace recognition in photos and spoken language translation have been offered by AWS, Google Cloud, and Azure Now offering AI-based platforms to almost any type of company, regardless of its size and technical sophisticationOther companies like Apple, IBM, Oracle, Salesforce, SAP, Alibaba and Baidu also have massive computing resources and talents required to build this AI utilityAI could dramatically multiply the size of the cloud marketWinners will control the OS of the future and become the most powerful companies in historySource: Peter Burrows, MIT Technology Review, Mar. 22, 2018
Slide14Source:
tefficient.com, THE SECRET BEHIND ELISA’S FINANCIALS, 2/4/2018
Slide15In 2008, DARPA issued a
challenge to researchers:Create a sophisticated, shoebox-size system that incorporates billions of transistors, weighs about 3 lbs, and requires a fraction of the energy needed by current computers—which migrates from recognition to perceptionBasically, a brain in a boxThe challenge triggered neuromorphic computing researchIBM TrueNorth ComputerIntel Loihi Test ChipIntel Neuromorphic Research Community (INRC)Neuromorphic Computing Challenge
Source: 1) HP Enterprise Labs, www.labs.hpe.com/next-next/brain
2) www.research.ibm.com/articles/brain-chip.shtml
3) newsroom.intel.com, Mar. 1, 2018
Slide16Open Issues of
Neuromorphic Computing:Basic building blocks and general architecturesConfiguring a neuromorphic device/tissueTraining (programming) a neuromorphic computerSystem softwareAppropriate applicationsComputing hardware development
Slide17Federal Vision for Future Computing[
White paper by DOE, NSF, DOD, NIST, IC (July 29, 2016)]Emerging computing architecture platforms, neuromorphic, quantum, etc.
Significantly enhancing performance while reducing energy consumption by
over 6 orders
of magnitude (from
MWatts
to Watts)
Intelligent big data sensor: autonomous and reprogrammable
Machine intelligence for scientific discovery
Cybersecurity
Slide18PCAST Report to the US President
Source: Report to the President: Ensuring Long-Term U.S. Leadership in Semiconductors, the President’s Council of Advisors on Science and Technology (PCAST), White House, Jan. 2017Semiconductors are essential to modern life
Creating new businesses and industries
Bringing massive benefits to American workers and consumers
Cutting-edge semiconductor technology is critical to US national security
The US should
Establish policies aimed at 1) developing and attracting talent, 2) funding basic research and development, 3) reforming corporate tax laws, and 4) reforming permitting practices
Establish a series of
moonshots
that would deliver
radical semiconductor advances
of much broader applicability, driving transformative innovation
Slide19Electronics Resurgence
Initiative (ERI)DARPA’s just earmarked $1.5B to fund HW projects through ERI that was announced in June 2017Goal: semiconductor innovation and circuit design upgrade in the USSome programs under the ERI umbrella:Intelligent Design of Electronic Assets (IDEA)EDA tools that learn, and are super productivePosh Open Source Hardware (POSH)Open source design and verification ecosystem for complicated SoCsDomain-Specific SoC (DSSoC)Software Defined Hardware (SDH)Three Dimensional Monolithic SoC (3DSoC)Foundations Required for Novel Compute (FRANC)Near Zero Power RF and Sensor Operations (N-ZERO)Partners: Major universities and companies
Source: DARPA 1
st
ERI
Summit, July 23-25, 2018, San Francisco
Slide20半導體產業的關鍵角色半導體
競爭有如核武競爭核武一:5G通訊晶片基頻:高通(1級),Intel(1級),三星(2),聯發科(2.5),海思(2.5),展訊(3)
射頻前端:
博通
、
Infineon
、
Kyocera
、
Murata
村田、
Qorvo
、高通、
Skyworks、 Taiyo Yuden 太陽誘
電
今年初
如果
博通買了高通,
Intel
準備出手買博通
核武二
:
AI
晶片
傳統
IC
公司:
Nvidia
(
1
)
,
Intel
(
1
)
,
IBM
(
1.5
)
,
Qualcomm
(NXP)
(
1.5
)
,
NXP
(
1.5
)
,
AMD
(
1.5
)
,三星
(
2
)
,海
思
(
2.5
)
,聯
發科
(
3
)
等
系統與雲端公司:
Google, Apple, Microsoft
, Tesla, Amazon, F
B
,
百度
,
阿
里巴巴等
核武三:車用
晶片
NXP, Infineon,
Renesas
, STM, TI, Bosch, ON
Semi, Microchip/Atmel, Toshiba,
Rohm
等
技術
(經濟)才
是
台灣
保
命的關鍵!
技術來自人才與研究
重
技術才會有人才,有人才才會有
投資
,有
未來