Dr. Wally Rhines
Dr. Walden C. Rhines (Wally), Cornami, Inc., presented on Shaping the Global Digital Future Through Secure Information Sharing, at the ongoing SEMI Technology Unites global summit.
Dr. Rhines said that one of the key parts has been collaboration to share information. The Moore’s Law was that transistor density doubles every 1-2 year. There was level playing field in the IP, thanks to broad AT&T’s licensing. There was free flow of people and ideas. The growth of transistor unit volume also led to sustainable growth. As the industry infrastructure evolved, the semiconductor industry also grew.
The learning curve for IC fab equipment was similar to semiconductors. Even EDA software costs had to conform. Free transfer of people led to international conferences and meetings, etc. There was MNC manufacturing at Europe, USA, Asia, and Japan. There were the EIAJ, SIA, SEMI, and other discussions. There was MITI involvement to encourage agreements.
Current challenges to global learning saw the drive for regional self-sufficiency and regional standards, etc. In export regulations, there was deja vu all over again. Countries like Taiwan became buyers of semiconductor manufacturing equipment. Re-activation of fabless AI companies received VC funding. It took off in 2017 with AI leading the growth. Today, the combined semiconductor market share of the 50 largest companies is less than the new companies.
Virtual meetings have made connecting easier. There has also been an increasing trend of semiconductor content in systems. There has also been progress through information sharing. Data has become the great value and the new oil. There are three approaches to share the value of data. Bring your data to the algorithm, push the algorithm to your data, and homomorphic encryption, which allows you to keep the data encrypted.
Fully homomorphic encryption (FHE) is the ultimate solution for cyber security. DARPA calls this the holy grail of cyber security. FHE provides the quantum secure encryption. It enables all types of arithmetic and logical computation, while data is encrypted. Effective machine learning also requires FHE. You can get training data privacy, input privacy, output privacy, etc. Dr. Rhines added that ML market for software will reach $20 billion by 2024.
FHE is also a very hard computational problem. However, semiconductor companies can achieve the solution. Massive parallelism is required at today’s clock rates. Minimize memory reads and writes. Most memory access must be local to the processor. You need software that can generate independent, executable streams of data and control. The chip hardware should scale linearly with the number of processor cores. Scalability is essential across multiple chips, boards and servers.
There are factors driving the adoption of FHE. Cloud security is the critical issue. The US Department of Defense (DoD) has taken a firm position supporting FHE. In fact, FHE has arrived just in time. Our global digital future depends upon collaboration. Free trade and free markets have powered the historical growth of semiconductor industry. With ‘confidential computing’, we can capitalize on the value of data, and massively impact the quality and productivity of the medical care and financial systems.
Chinese American Semiconductor Professional Association (CASPA) organized its annual conference: Next Wave of Semiconductor Innovation, in the USA.
Delivering the keynote titled: Next Wave of Semiconductor Innovation, Dr. Walden C. Rhines (Wally), President and CEO, Cornami, and CEO Emeritus of Mentor, a Siemens Business, looked at some of the big possibilities that will happen.
Waves of new product categories have been driving the electronics industry revenue. Wireless communications currently generates 40 percent of growth. The next wave of growth will be driven by data. It is all about large data. Managing and analyzing large volumes of data will be key. AI/ML will build ML models. IoT will bring computation close to the point of data collection. 5G is also there.
AI had happened back in July 1986, a technology that arrived before its time! Back then, there was a lack of big data to analyze. Computers were not as powerful. However, all the limitations are going away, today. This will usher in an era, where data can become the new oil.
Data new oil
For example, the aircraft engine manufacturers could give away the engines for free, and charge for the support and maintenance. There are three things needed to make data the new oil. We are collecting a lot more data than we can analyze. Video surveillance alone generates petabytes per day. There are a myriad of new apps for solid-state imagers.
The IoT intelligent edge devices require technology integration. There are complex, but extremely cost-sensitive design. Each packaging era also brings disruptive change. Chiplets are becoming the next customization approach. The value of IoT is in the information collection and analysis.
Eg., Google has designed a chip to collect medical information. It can potentially track 100s of bio markers. Google is in the information business. Tesla replaced general-purpose GPU with custom chips. System companies continue to be the fastest-growing foundry customers, growing at a 5-year CAGR of +35 percent. Leading-edge systems companies are becoming the new SoC designers and manufacturers.
Analysis of data is next! Domain-specific processes have been there. Traditional Von Neumann computer architectures have been efficient for pattern recognition. Neural networks are a fundamental building block for AI-related ML. Fabless AI companies have received venture capital funding. Investment has remained heavy since 2018, at $2 billion.
Start-ups are also dominated by domain specific architectures. A large category is in vision/facial recognition, with 45 companies. SenseTime is a leader in facial recognition. There is the Google Tensor processor of ML training and inference. Edge computing has 36 companies. This business will grow even faster. Intelligence flows downhill over time!
We also need to protect the data! You don’t want to make personal data available on computers. There are three approaches to privacy computing. Bring your data to the algorithm. Push the algorithm to your data. Trust the code in your data center. You also add homomorphic encryption, and trust no one. Secure the data, and not, the data center.
Focus on FHE
DARPA calls all of this the holy grail of data security. Privacy preserving ML preserves investments. Trained ML models are valuable. Eg. GPT-3! There can be ML using encrypted data. Fully homomorphic encryption (FHE) makes ML training possible using encrypted data. FHE is the holy grail of computing. FHE is a very hard computational problem. Computational fabric enables real-time FHE. There are attempts to implement FHE. A top fintech company applied FHE in the AWS Cloud with medium-sized encrypted databases of 100 million entries.
There are requirements for achieving 106x performance at lower power. There is massive parallelism at today’s clock rates. Factors driving the adoption of FHE are cloud security. The DoD has taken a firm position supporting FHE. Hundreds of companies have homomorphic encryption programs already, by 2025. Quantum computers will also break the current Internet security. This could happen between 2023-2030.
The next generation of semiconductor growth will be based on making data the new oil. ICs are capturing an increasing share of the electronics system product value. The IC content is increasing the share of value.
What is FHE?
FHE has long been described as transformative for cloud security. The algorithm was developed to enable computing on encrypted data sets, keeping the underlying data secure. It is a game-changer in cloud computing as it allows for extracting valuable data analytics without ever decrypting the data to expose the underlying data, whether it is sensitive intellectual property (IP), financial information, personally identifiable information (PII), intelligence insight, or beyond.
FHE is a significant and critical technology for enterprise, industry and governments given the cost and damage associated with the growing number of cybercrimes and breaches that is at the trillions-of-dollars level. However, the challenge to date has been that it is too computationally expensive to be commercially practical.
Earlier, Dr. Song Xue, Chairman and President, CASPA, welcomed the audience. He looked at the accomplishments in the year gone by. There have been economic uncertainties, Covid-19, and China-US trade war. CASPA formed a special group for support to Wuhan and China. They also sent PPEs to different hospitals. The CASPA annual conference was also held. The CASPA 2020 Summer Symposium was also held this year. BIMS IC Conference also had a delegation from CASPA. Looking forward, we know there will be many challenges, and will overcome those.
Kansen Chu, Assembly Member, California State Assembly, said CASPA provides great opportunities for its members. A certificate was also presented to CASPA. He added that the industry recovery will take some time.
Brandon Wang, Chairman, Board of Advisors, CASPA, and VP, Synopsys, said that CASPA has done very well to host many events this year. 2020 has been a year of shock and unexpected happenings. Moving forward, there will be new norms and a new vision for 2021.
Next, Dr. Xiaodong Zhang was selected Chairman and President, CASPA, 2021. He shared the magic of semiconductors. Semiconductors have been the backbone of the technology industry. We are currently engulfed by Covid-19. We don’t need to panic, though. We are looking at the opportunities being provided.
There is strong data center and PC demand given WFH. There is strong demand for smart appliances, as well. The future vision is where we can grow the connections, do communications, and advance our careers. The last area is college program, where professors also share their research with members.
The 2020 IEEE International Reliability Physics Symposium (IRPS), the industry’s premier technical conference for engineers and scientists, to present the latest original research in microelectronics reliability, will be held in Dallas, TX, USA, from March 29 – April 2, 2020 at the Hilton DFW Lakes Executive Conference Center.
The symposium will feature a number of special focus sessions, highlighting novel and emerging areas of electronic reliability, as well as topics relating to conventional semiconductor, integrated circuit, and microelectronic assembly reliability.
It will highlight the latest research in reliability for semiconductor devices, microelectronic systems, and advanced technologies.
There are topics such as:
- Wide bandgap semiconductors – reliability topics on SiC, GaN, and gallium oxide devices
- Neuromorphic computing reliability – memory and design architectures
- Circuit reliability and aging – EDA tools, compact modeling, and aging-aware designs
- Reliability of RF/mmW/5G devices – CMOS, SiGe, BiCMOS, SOI, and GaAs.
There are four special focus sessions at IRPS that will combine invited and accepted papers, tutorial sessions, workshops and poster presentations in these emerging microelectronics reliability topics:
Silicon carbide device reliability – SiC is gaining maturity in high power/high voltage semiconductor applications, surpassing GaN in some cases. This focus session includes three invited talks addressing defects in SiC, ruggedness under extreme conditions, and development of new SiC standards.
Neuromorphic computing reliability – As CMOS scaling slows down, improvements in von Neumann computing become more difficult. This focus session features three invited talks on the reliability aspects of various memory devices used for neuromorphic computing.
IC reliability and aging – Performance requirements of highly-scaled IC designs are colliding with design margins needed to account for device degradation due to TDDB, BTI, HCI, and EM. Four invited talks in this session highlight the modeling of aging, in-field aging data collection/analysis, and design for aging.
Reliability in RF/mmW/5G – Mobile communications for 5G networks and autonomous vehicles will require devices operating in the RF and millimeter-wave regime. This focus session includes three invited talks encompassing GaN, Si, and III-V device technology for high speed applications.
Dr. Walden Rhines, Mentor, will be delivering a plenary keynote, titled: Reliability Drives Semiconductor Industry Evolution.
The attention to reliability issues has evolved from a focus on burn-in and rigorous testing to designed-in reliability and evolving methods for modeling physical failure mechanisms, intelligent verification of layouts and machine learning.
The other plenary presenters are going to be Dr. Mike Mayberry, Intel, Dr. Oliver Häberlen, Infineon, and Dr. Alessandro Piovaccari, Silicon Labs.
Two full days of 90-minute tutorial sessions have been designed to provide a comprehensive overview of the reliability topics. There will also be interactive evening workshops covering 10 different reliability focus areas. A supplier exhibition will be held in conjunction with the technical program.
Predicting Semiconductor Business Trends After Moore’s Law is a book written by Dr. Walden C. Rhines, CEO, Rhines Consultants and CEO Emeritus, Siemens PLM Software and Mentor, a Siemens Business. It is a SemiWiki project.
In the book, he begins with how the semiconductor learning curve provides a roadmap. According to him, Texas Instruments’ (TI) unique approach for semiconductors lay in the use of the learning curve to drive a pricing strategy early in the life of a new component.
While Moore’s Law is becoming obsolete, the learning curve will never be. Instead, the cumulative number of transistors produced will stop moving so quickly to the right on the logarithmic scale. Then, the prices will not decrease as rapidly, as they have in the past. The visible effect of improved learning will diminish.
A graph shows how the cumulative unit volume of transistors used in memory components is increasing much faster than unit volume of transistors in other types of chips.
He says that if the EDA industry doesn’t keep its learning curve parallel to the semiconductor industry learning curve, the cost of EDA software as a percent of semiconductor revenue would increase. There would have to be cost reductions elsewhere in the semiconductor supply chain to offset it.
About Moore’s Law
Dr. Rhines goes on to say that ‘Moore’s Law’ has been extrapolated for more than 50 years. It is not a ‘law’. It is an empirical observation that became self-fulfilling after some adjustments. He cites revisions by Gordon Moore to the Moore’s Law, at least twice, in 1975 and 1997. And, later, in 2003. These repeated revisions affirm that “Moore’s Law” was not actually a law of nature, but an interesting, if temporary, phenomenon.
Dr. Rhines says that today, many people worry that the inevitable end of Moore’s Law will leave us with a stagnant semiconductor industry, with no guideposts to drive new silicon technology directions. Fortunately, these people need not worry. The learning curve is valid forever (when measured in constant currency, corrected for governmentally-induced inflation) as long as free market economics prevail, i.e., negligible trade barriers, no regulatory price controls, etc.
In 1825, Benjamin Gompertz proposed a mathematical model for time series that looks like an ‘S-curve’. The Gompertz Curve has been used for a variety of time dependent models, including the growth of tumors, population growth and financial market evolution.
The actual rate of growth of shipments of silicon transistors is predicted to increase until about 2038. At that time, the Gompertz Curve suggests that the increase in the rate of growth will become zero, and the rate of increase will be less each year, until we reach saturation, sometime in the 2050 or 2060 timeframe. By then, we should have developed lots of alternatives.
In another chapter, there is a graph detailing how the top 50 semiconductor companies’ share of the market has decreased 10 points in 10 years. He adds that it’s difficult for semiconductor companies to re-invent themselves as new growth markets emerge. Large semiconductor companies tend to grow at about the overall semiconductor market average growth rate, while the new entrants grow much faster, albeit from a smaller revenue base. Gradually, these small companies climb the ranks on their way to top 10.
Elsewhere, he talks about how EDA has evolved to an extent that the complex chips with tens of billions of transistors frequently produce first pass functional prototypes from the manufacturer.
International semiconductor competition
Later, he touches upon the international semiconductor competition. Semiconductor industry evolution was largely a US phenomenon. Japan became a significant competitor, especially in the late 1970s, and early 1980s. Then came Korea. Next, there was the evolution of worldwide leadership in the silicon foundry business by Taiwan, which is truly remarkable. TSMC also recognized the value of being a dedicated foundry, with no products of its own to compete with its customers.
Despite China’s rise as the world’s largest assembler of consumer electronic equipment, the Chinese semiconductor industry has evolved slowly. The largest Chinese semiconductor foundry, SMIC, is said to be two technology nodes behind TSMC in manufacturing capability, as of 2019. The Chinese government is dedicated to changing this situation.
China has done the expected. They have focused upon developing non-US capabilities for all their components. Since China buys more than 50% of all semiconductor components in the world, and uses more than 15% of the world’s semiconductor supply in equipment designed by Chinese companies, this is now a big problem for the US semiconductor industry. It is probably not reversible.
While China’s direction is not likely to change, we still have the possibility of convincing the rest of the world that the US can be treated as a reliable supplier. Hopefully, there will be policies articulated by the US that convey that confidence, and restore the US position as a leader in free trade.
AI, edge computing and 5G
Towards the end of this engrossing book, Dr. Rhines touches upon artificial intelligence (AI). He says, and I agree, that AI is not a new technology. Here is the cover of High Technology magazine in July 1986. Dr. Rhines is the person on the left, and George Heilmeier, former head of DARPA, is on the right. “We tried hard in the 1980s, but the infrastructure had not developed to a level where AI would provide profitable opportunities,” he added.
Today we have overcome all these limitations. AI and ML have taken on a life of their own. They have become limited, however, by the processing power available. Traditional von Neuman general purpose computing architectures are inadequate to handle the complex AI algorithms. The result: a new generation of computer architecture is evolving.
On edge computing, he says that the edge nodes will require mixed technologies. Simulating digital logic connected to analog, RF and other technologies is not easy. A whole new family of EDA tools is required.
One of the great opportunities for the semiconductor industry is the increased number of base stations required to support the infrastructure of 5G and the larger number of antennas in a phone. It’s likely that about a dozen companies will lead the way in supplying the complex image processing subsystems required for autonomous vehicles.
May I take this opportunity to thank Dr. Wally Rhines for sending me a copy of this superb book. Friends, you can read this too on SemiWiki. It’s really a brilliant read! Enjoy! 🙂
You’re a good soldier, choosing your battles
Pick yourself up and dust yourself off and back in the saddle
You’re on the front line, everyone’s watching
You know it’s serious we’re getting closer, this isn’t over
The pressure is on, you feel it, but you’ve got it all, believe it
When you fall get up oh oh, and if you fall get up oh oh
Tsamina mina zangalewa, ’cause this is Africa
Tsamina mina eh eh, Waka waka eh eh
Tsamina mina zangalewa, this time for Africa! 😉
— By Shakira, from World Cup Soccer, 2010! 😉
We have come to the end of a very interesting decade! While it was not so engrossing as the 2000s, there were several developments worth noting. But first, let me take you all back to March 15, 2016!
Around 2.30am, I was in the bathroom. All of a sudden, my legs simply gave way! I didn’t know what was happening to me. Also, there was a sudden increase of chest pain! I clung on to the bathroom door, and somehow crawled to my bedroom. There, I tried to wake up my wife! By the time she was up, I was lying down on the floor, sweating heavily, and blacking out! Mind you, I never drink!
My wife and brother rushed me to Sodhi Nursing Home, where the doctor diagnosed me with a severe heart attack. He recommended that I be immediately taken to Action Balaji Hospital. There, the doctors took one look at me, and rushed me to the operation theater. I was later told that I had a massive heart attack, with 99 percent blockage in my veins. I don’t even know what the doctors did, but here, I am before you, presenting my state! This is a trend, I never even imagined, would happen to me!
Given here are some of the global technology trends and happenings that shook the world during the last decade:
Mobile Internet, Bluetooth and Wi-Fi
Back in late 2000, at the ITU World Telecom event in Hong Kong, the first mobile phones with Internet browsing were being touted. Back then, mobile Internet was all the rage! As, were 3G and Bluetooth! This was the 3G technology based on W-CDMA and also, TD-SCDMA. Those were also the days when ‘WAP is CRAP’ made more headlines, and bore the brunt of many ‘telecom jokes’! Today, we can’t even imagine a life without the mobile Internet! And, we are greatly bothered if we can’t access a page on our mobiles!!
In early 2002, I wrote an article for Electronics Business News Asia (EBN Asia), Singapore, on Bluetooth, which was still trying to find its bearings. I can’t locate that article anymore! Some of the comments are worth remembering. One comment was whether Bluetooth and Wi-Fi could co-exist! Today, the world is into launching Wi-Fi 6 and Bluetooth 5.1!
March 2011, we saw the Japanese earthquake – The Japanese earthquake and tsunami stunned the global electronics and semiconductor industries!
Tsunami and earthquake
The preliminary assessment of Texas Instruments’ manufacturing sites in Japan revealed that the fab in Miho suffered substantial damage during the earthquake. Teams are working to reinstate production in stages, reaching full production in mid-July. TI’s fab in Aizu-Wakamatsu was damaged, but was being re-started with full production estimated by mid-April. TI’s third fab in Hiji was undamaged and running at normal capacity.
Sony Group Operations were said to have been affected by the Pacific coast of Tohoku earthquake, tsunami and related power outages. For Elpida, the Hiroshima Plant suffered little impact as it is located in Hiroshima in the southwest of Japan, However, the Akita Elpida memory plant is not in operation as of the time of the announcement due to power shut down caused by the earthquake, and it is hoped that normal business will resume when the power returns.
Iwate Toshiba Electronics did not report any casualties, but as of March 15, there was power lost, with limited partial recovery to start from March 13. As of March 15, 12:00pm, seven factories out of 22 of the Renesas Group’s factories in Japan temporarily shut down production.
The Shin-Etsu group reported that as of 1pm, March 15 (Japan Time), necessary inspections were carried out at Shin-Etsu Chemical Kashima Plant (Kamisu, Ibaraki Prefecture) and Shin-Etsu Handotai Shirakawa Plant (Nishigo Village, Fukushima Prefecture), both of which were out of operations.
Mitsui Chemicals Group reported the effects of the Kanto-Tohoku earthquake on its operations. The operations at the Kashima Works (Kamisu City, Ibaraki Prefecture), was suspended since the earthquake. Operations were resumed after assessment of damage by the earthquake and tsunami.
At its Ichihara Works (Ichihara, Chiba Prefecture), production at the ethylene plants was according to schedule. The operations at Mitsui DuPont Polychemicals and Chiba Phenol plants were suspended since the earthquake.
At the Mobara Branch Factory (Mobara City, Chiba Prefecture), operations at acrylamide and paint toner binder resin plants have been suspended since the earthquake. After assessing effect of scheduled “rolling” blackout, operations were resumed.
USB 3.0 also became widely available, while 22nm chips entered mass production. Consumer-level robotics were also booming.
Birth of EVA
In May 2011, the Embedded Vision Alliance was born! Over 15 leading technology companies, came together in Oakland, USA, to ‘speed up the adoption of computer vision capabilities in electronic products.’ BDTI, Xilinx, and IMS Research initiated the EVA, and were joined by Analog Devices, Apical, Avnet Electronics, CEVA, CogniVue, Freescale, NVIDIA, National Instruments, Texas Instruments, Tokyo Electron Device, MathWorks, Ximea, and XMOS as the founding members.
Still in June 2011, June 8 happened to be World IPv6 Day. Google, Facebook, Yahoo!, Akamai and Limelight Networks were among some of the major global organizations offering content over IPv6 networks on a 24-hour test flight! World IPv6 Day’s goal is to motivate organizations — ISPs, hardware vendors, OS vendors, web companies, etc., to prepare their services for IPv6, as IPv4 addresses ran out! IPv6 was designed to succeed the IPv4.
End of Harry Potter, Steve Jobs
In July 2011, we saw the end of the spectacular Harry Potter movies! Right from the time Harry confronts Helena Ravenclaw or the ‘Grey Lady’, the Death Eaters attacking Hogwarts, the very brave resistance put up by the school inhabitants, led by Prof. Minerva McGonagall, the tragic death of Severus Snape at the hands of Voldemort and his pet snake, Nagini, and Snape’s final meeting with Harry, following which, Harry views Snape’s pensieve and learns about his love for Lily Potter, up to the time Harry enters the Forbidden Forest to meet his death! Or, was it Harry, or Voldemort, who dies? It’s all breath-taking!
October 2011, Steve Jobs, the master of the game, is gone! I first had a look at the Apple Mac, while at SBP Consultants & Engineers, back in 1988. I was surprised to find a computer that could do desktop publishing so well! By then, Jobs had gone out of Apple, fired by John Sculley, then Apple’s CEO, sometime in 1985. Jobs returned to Apple in 1996, a time when he had floated PIXAR and NeXT — the company Apple eventually bought, and returned Jobs to Apple. The rest is history!Read the rest of this entry »
Dr. Walden C. Rhines, CEO Emeritus, Mentor, a Siemens Business, is in Bangalore, India, at Mentor U2U India, 2019. He delivered the keynote during the day-long conference.
Why is the semiconductor industry design activity accelerating? Dr. Rhines said that the acceleration of semiconductor revenue growth, in terms of the annual growth of ICs has been 2.8 percent in a 5-year CAGR from 2011-2016. It was 22.2 percent in 2017 and 2018E (estimated) is 15.5 percent, according to VLSI Research, Jan. 2019. There has been an acceleration in R&D investment, with semiconductor R&D spending touching 7.5 percent in 2018.
Memory prime driver
Fabless semiconductor venture capital investment by year, all rounds, has been $3,118 million in 2018 YTD (year to date). Memory was the primary driver of semiconductor growth in 2017. Memory is now almost 40 percent of IC revenue vs. 27 percent in 2016.
Memory average selling prices (ASPs) increased dramatically in 2017, and in H1 of 2018. Memory prices peaked in Q1, but are expected to decline in 2019. Negative memory ASP year-to-year growth in Q418 and forecast for next four quarters, with MOS logic at 6 percent, MOS memory at 57 percent, and analog devices at -8 percent. Memory unit volume now needs to grow faster than logic!
Memory now dominates transistors manufactured. 3D NAND allows memory to scale faster than logic or SoCs. Despite 55 percent revenue growth in 2017, the memory unit volume growth is below the long-term trend line. Will non-memory growth remain strong when memory ASPs decline?
New companies in semicon design
New companies are now entering the world of semiconductor design. Amazon has just become a chip maker. Bosch has opened a billion-dollar wafer fab. Google has built its first chip for machine learning. Facebook plans to build its own chips as part of hardware push.
The Internet has greater than 22 billion connections today. Of these, 81 percent of the connections are things. Things are growing at 15 percent CAGR, wireless mobile at 11 percent CAGR, and stationary/desktop at 5 percent CAGR.
The Industrial Internet and connected cities are 79 percent of the market today. Connected vehicles are showing the fastest growth. Connected homes are growing at CAGR 18 percent, connected vehicles at CAGR 23 percent, wearables at CAGR 21 percent, Industrial Internet at CAGR 21 percent, and
connected cities at CAGR 7 percent, as of the end of 2018.
Chinese investment growing
Chinese investment is now moving from manufacturing to design. The Chinese government’s incentives for semiconductor investment have been rising. The government-backed China IC Investment Fund is worth US$ 47.4 billion (Yuan 300 billion).
The China semiconductor initiative has accelerated new startup formation. IC design enterprises have gone up from 715 in 2015, to 1,380 in 2017. China’s fabless companies have also become much larger, from 479 enterprises in 2006 to 715 in 2015. And, this number is only rising.
China fabless semiconductor companies by market segment has 935 unique companies – who can be in multiple sub-segments. There are 961 companies in power devices, 267 in analog devices, 266 in MEMS/sensors, 260 companies are fab owners, while 209 are in RF devices. Interface chips have 179 companies, memory devices have 135, full custom devices have 124. Video compression has 82 companies, while vision processing/AI/ML has 35 and there are 34 photonics suppliers. All other categories (semi and system) have 1,325 companies.
Neuromorphic computing is said to be for the next wave of automation. The
Evolution of non-von Neumann computer architectures will improve processing speed, reduce power and integrate more memory.
Traditional Von Neumann computer architectures are not efficient for pattern recognition. Computer architectures are a long way from human brain pattern recognition and power dissipation. Large number of computer cycles are required to perform the same level of pattern recognition as the human brain.
Neural networks are a fundamental building block for AI-related machine learning. Today’s artificial intelligence scenario looks like this: in 2017, more than 300 million smartphones shipped with some form of neural-networking capabilities. In 2018, 800,000 AI accelerators were shipped to data centers. Every day, 700 million people use some form of smart personal assistant like an Amazon Echo or Apple’s Siri.
India’s strengths in AI
TheIndian industrial companies are among the early adopters of AI. Early adopters are defined by BCG as businesses that have fully implemented more than one AI use case in multiple industrial operations areas. India ranks third, at 19 percent, behind the USA, at 25 percent, and China, at 23 percent, respectively.
India’s design activity is accelerated by investment activity. There is the Semiconductor Fabless Accelerator Lab (SFAL), which the IESA launched in December 2018. It has investment by the Government of Karnataka. SFAL plans to accelerate 20 startups in the next three years, and 50 in next five years. It will support the existing fabless companies with a goal of at least 2-3 products out of the accelerator over the next two years.
There is also the Fabless Chip Design Incubator (FabCI), launched in 2018 by IIT Hyderabad. It is an incubator for fabless chip design startups. Funded by the Ministry of Electronics and IT, technology partners are Cadence Design and Mentor. Its goal is to incubate at least 50 ‘Make-in-India’ chip design companies.
IITs are supporting the AI education. IIT Hyderabad offers a Master’s degree in AI. It is adding a Bachelor’s program in AI during the 2019-20 academic year. IIT Kharagpur has a six-month AI certificate course. IIT Madras has the Robert Bosch Center for Data Science and Artificial Intelligence. India has a strong AI skillset, and ranks third, behind the USA and China.
There are two key requirements for brain-like pattern recognition — memory improvements and processor architecture improvements. In memory improvements, there are increased capacity, hierarchical memory, memory cell connectivity, and invariant memory. Processor architecture improvements have parallelism, error tolerance, continuous feedback and integration with memory.
Major wave of new ‘domain-specific’ architectures
There is the end of the Moore’s Law and faster, general-purpose computing, and a new golden age. Software-centric, modern scripting languages are interpreted, dynamically typed and re-use is encouraged. Hardware-centric, the only path left is domain-specific architectures. Just do a few tasks, but extremely well. The combination of software and hardware gives domain-specific languages and architectures.
Startups dominated by application specific architectures in the worldwide fabless company VC funding show AI and ML topping at $1,834 million between 2012-18. AI is NOT a new technology. It is a technology that arrived before its time, in 1986. (Dr. Rhines is among the heavyweights involved).
Reasons for AI adoption delay in the 1980s were lack of big data to analyze, as there was no Internet or IoT to collect sizable data sets. There was limited computing power, and limitation of traditional computer chip architectures. There was a need for more advanced algorithms. Besides, there was lack of ‘killer’ applications to make money! What’s different today? All of these limitations are going away.
24 fabless AI companies received VC funding in 2018. Some are AIMotive Gmbh, Beijing Intengine Tech, Hailo Technologies, Mythic, Syntiant and Xanadu Quantum Tech. Also, the early round China fabless funding passes the USA.
If you look at the domain-specific AI/deep learning controllers, there are 39 for vision/facial recognition, 9 for voice/speech/sound/pattern recognition. 17 are for autonomous driving/ADAS, three each for disease diagnosis AI and optical computing AI, 2 for smell/odor recognition, 23 for data center/cloud AI/HPC, 6 for unknown/stealth mode, 3 for space/military apps, 2 for cryptocurrency, 21 for edge computing, 4 for robotics/motion control/collision avoidance, 8 for deep learning training, and 1 for intelligent wireless control. Pattern recognition dominates the new AI designs.
Microsoft has developed custom deep neural network chip for HoloLens. HoloLens is an untethered mixed-reality device. It has local compute for low-latency (fully battery powered). It includes custom holographic processing unit (HPU), custom AI coprocessor, custom time-of-flight depth camera, four gray-scale cameras, IMU, other sensors, and infrared camera.
EMOSHAPE has developed an emotion processor unit (EPU), which has the capacity to feel, perceive or experience subjectively. Emoshape has developed its own CPU optimized to handle emotional data. The technology has the potential to change computer games, virtual reality and augmented reality applications, said Roberta Cozza, research director, Gartner.Read the rest of this entry »
Happy new year, to all of you. 🙂 And, it gets even better, having a discussion with Dr. Walden C. Rhines, CEO and Chairman of the Board of Directors of Mentor, A Siemens Company, on the global semiconductor industry trends for the year 2019.
Semiconductor industry in 2018, and 2019
First, I needed to know how did the global semiconductor industry performed last year? And, what is the way forward in 2019.
Dr. Wally Rhines said: “2018 was another strong growth year for the global semiconductor. IC bookings for the first 10 months remain above 2017 levels and silicon area shipments for the last six quarters have also been above the trends line, with fourth quarter YoY growth 10 percent. And, IC revenues overall continue to have strong double-digit growth for 2018, with fourth quarter YoY growth of nearly 23 percent.
“However, analysts are expecting much more modest growth in 2019. Individual analyst predictions for growth in 2019 vary from -2 to +8 percent, with the average forecasts at +4.4 percent.
“Much of this is due to the softening memory market, along with concerns about tariffs, inflation and global trade war. While the rest of the IC business has been relatively strong with Samsung and Intel noting solid demand for ICs for servers and PCs, sentiment by senior managers of semiconductor companies is near a record low level. So, I’m not expecting much growth, if any, in 2019 and more likely a decline.
EDA in 2019
On the same note, how is the global EDA industry performing, and what’s the path in 2019?
He said: “Revenue growth of the EDA industry continues to be remarkably strong, fueled by new entrants into the IC design world, like networking companies (e.g. Google, Facebook, Amazon, Alibaba, etc.) and automotive system and Tier1 companies, as well as a plethora of new AI-driven fabless semiconductor start-ups. Design activity precedes semiconductor revenue growth so it would not be surprising to continue to see strong EDA company performance even with a weak semiconductor market in 2019.
“EDA venture funding has rebounded, reaching a 6-year high of $16.5M showing a renewed confidence in the future of EDA. The major companies all have sighted better than expected results. On the semiconductor side of EDA there seem to be more technology challenges than the industry has faced in a long time.
“Some of those include new compute architectures, the emergence of photonics, increased lithographic complexities involving EUV and other techniques, new and more complex packaging, massive increases in data, and the multiplication of sources of design data (often created according to differing standards).
“The challenges on the system side of EDA are multiplying as expected. It is becoming more difficult to be at the leading edge when designing end-products in silos. Embedded software, mechanical, PCB, packaging, electrical interconnect, networking (access to the intranet) and security are just a few of the domains that need to work closer together in a more integrated manner. The increasing complexity is also making each of the domains more challenging. This all pushes new materials and methodologies into each of the domains listed above.”
Five trends in semicon for 2019
I wanted to find out about the top five trends in semicon for 2019.
He said: “The top five semiconductor technology trends include:
* the ongoing ramp of next-generation technologies, led by Machine Learning, Artificial Intelligence and cloud, and SaaS demand on the datacenter,
* the roll-out of IoT – especially in manufacturing,
* 5G development,
* computing on the edge, and
*the increasing semiconductor content within electrical devices.”
According to an IC Insights report, the 47 percent full-year 2017 jump in the price-per-bit of DRAM was the largest annual increase since 1978, surpassing the previous high of 45 percent registered 30 years ago in 1988! This sounds interesting!
Are the rising DRAM prices aiding startup Chinese competitors? Are major DRAM suppliers somehow stunting global DRAM demand?
Dr. Walden C. Rhines, president and CEO, Mentor Graphics, a Siemens Business, said: “The DRAM business has always gone through cycles of imbalance between supply and demand. Growth of demand in the last 18 months has been stronger than growth of supply.
“Substantial investments in 2017 by the MOS (metal-oxide semiconductor) memory producers, as well as the addition of China to the supply chain, will correct this imbalance late this year or, at the latest, early next year.”
The DRAM price-per-Gb has been on a steep rise. To this, Dr. Rhines said: “It is a commodity, although there are many types of specialty DRAMs emerging. Because DRAMs are viewed by customers as a near-commodity, the price is heavily influenced by the availability of supply. Supply has been very tight during the last 18 months.
Malcolm Penn, chairman and CEO, Future Horizons, UK, added, “This is supply and demand, pure text-book economics.”
Are the rising DRAM prices opening the door for startup Chinese competitors?
Dr. Rhines noted: “Chinese competitors made their decision to invest in DRAM capacity long before the recent strengthening of demand in the balance of supply and demand. Of course, higher, or stable, pricing may make it easier for new producers to absorb the costs of ramping up new capacity and developing experience with a new technology.”
Malcolm Penn agreed: “Potentially yes, and to anyone else. Coca Cola were contemplating building DRAMs in the 1990s. DSRAM market boom, again, pure text-book economics. Whether or not they succeed is an entirely different matter. If the Chinese do enter the market, can they then survive the inevitable downturn and cycles? That remains to be seen!”
Can the startup Chinese DRAM producers field any competitive product soon? Dr. Rhines noted: “They probably can. But, they will have to develop a production base of “learning” to reduce cost, improve yields and maybe even reliability. This will take some time.”
Penn added: “Technically (i.e., meeting the spec), probably, yes. Reliability, probably no, for the Tier 1 customers (that will take several years to build up the production experience). Cost, definitely not!
“Their small fab scale and late learning curve start means that their die cost will be sizably higher than those of Samsung and SKH, and also Micron. Plus, their yields will be lower. Then, there’s the deep cash pockets issue to fund these ongoing cost disadvantages.”
In a separate situation, some 300mm fabs closing, for example, ProMOS. Dr. Rhines said: “It’s because of an imbalance of supply and demand for the products they make, thus limiting their profitability. It could also be because they don’t see an adequate investment return from the expensive new capacity investments, and therefore, find it more attractive to phase out some of their existing capacity.”
Malcolm Penn felt that the fabs were too old and technically obsolete.
Finally, are there more IC companies making transition to fab-lite or fabless business model?
Penn noted: “There’s no-one left to change! Everyone’s now fablite or fabless, except for Intel and Samsung (logic) and the memory manufacturers.”
Dr. Rhines said: “Based upon the growth of foundry revenue vs. total semiconductor revenue growth, there must be a continuing transition of capacity away from IDMs toward foundries. In addition, IDMs like Samsung are finding it economic to build the foundry business to increase the volume base of products that utilize their technology and capital investment.”
Here is the concluding part of my discussion with Dr. Walden Rhines, chairman and CEO, Mentor, A Siemens Company.
Has the PSS been formally released? What are its implications?
Dr. Rhines said: “Accellera released an Early Adopter spec for public review at DAC in June, 2017 and is currently working on completing our work in preparation for a 1.0 release in 2018. Accellera plans to have a “1.0 Preview” version available in February, 2018 (@DVCon US) for another 30-day public review period. Then, they will do one more cleanup pass, and submit to the Accellera Board for approval in May 2018.
“The expectation is that the Board will approve the Portable Stimulus Standard 1.0 version in June, 2018, prior to the DAC. Mentor plans to have Questa inFact fully updated by then, to fully support the new standard when it comes out.
“As for the implications, we expect the Portable Stimulus standard to be the next advancement in abstraction and productivity for SoC verification. It is not expected to replace UVM, but rather be complementary to UVM to improve coverage closure, verification efficiency, and effectiveness at the block level.
“The ability to re-use the verification intent expressed in PSS from a block-level UVM environment to a software-driven, embedded-processor SoC environment, on multiple platforms (simulation, emulation, FPGA prototyping, etc.), will provide a quantum leap in productivity.
“Since the Portable Stimulus specifications are declarative, tools can fully analyze the verification-intent description at the system level and generate multiple correct-by-construction implementations of use case tests, on multiple platforms, from a single specification without requiring the verification team to rewrite the tests in UVM for the blocks and C for the system.
By the way, are the semiconductor/EDA companies re-looking at designs, rather than analyze more than 500,000 defective parts every day to identify design and process problems? If yes, how?
He said: “With today’s increased design complexity – they do both – re-look at designs before manufacturing and analyze afterwards. The complexity of today’s designs and manufacturing process requires multiple approaches to achieve high yields in each new node that is rolled out.
“Design for manufacturing and for yield are a must. However, the knowledge of the specific design practices that need to be followed for a new node is developed in multiple stages: in pre-silicon, test chips, first production design and when chips reach high production volume.
* Pre-silicon: Simulation models are used for initial design rules. Many assumptions are made and care must be taken to balance the benefit with potential overdesign for a process that will mature over time.
* Test chips: Early test chips try to mimic the major features of a real design, however, limited complexity and volume means some design rules can’t be discovered at this stage.
* First production design: Additional complexity of a real design and increased volumes expose more issues that need to be fed back to design for future revisions or the next design on a node.
* High production volume and additional designs introduced: High production volume and each subsequent design can benefit from the learnings at the previous stages. Many issues during this phase are resolved with process improvements, but continuous learning still remains key.
“The challenge is not eliminating the later learning phases, as this will never go away. Rather, the challenge is for the industry to maximize the learning at each phase and establish a continuous improvement cycle in design to take advantage of the knowledge gained. This is the foundational idea in closed-loop DFM, which is a process to maximize the design for manufacturing benefit throughout all phases.
Let’s also look at verification. What is the latest regarding coverage and power across all the aspects of verification?
Dr. Rhines added: “Actually, the recent trends have expanded to multiple concerns that cut across all aspects of verification, beyond coverage and power, such as security and safety. One driving force behind these trends is the convergence of computing, networking, and communications technologies. This is driving new markets, such as the Internet-of-things (IoT) ecosystem and automotive.
“A common theme across these emerging systems is the need for security, safety, and low power–whether you are talking about IoT edge devices or high-availability systems in the cloud. These new challenges have opened innovation opportunities, enabling us to rethink the way we approach verification. For example, concerning coverage, new statistical metrics have emerged providing deep system-level analysis capabilities that leverage data analytics techniques. This insight has become essential for system-level performance analysis.
Read the rest of this entry »