Delivering his keynote, Dr. Walden C Rhines, CEO Emeritus, Mentor – a Siemens Business, said that AI and ML are very popular topics nowadays. AI dwarfs anything else that we had in recent history. There is a lot of buzz around AI. What effect will AI and ML have? AI and ML will drive the new class of chips and architectures in the coming days.
In 2017, VC-funded fabless semiconductor companies invested $1.4bn, and soared to $3bn in 2018. The round 1 funding has been very high. It is a big opportunity for the semiconductor industry, and also in India.
Be aware that AI didn’t just turn up in H2 of 2017. It has been around for a long time. In 1986, AI turned up. But, there was lack of big data to analyze. We also didn’t have the Internet. There was limited computing power to do things with the data and a need for more advanced algorithms. The single biggest thing was the lack of any single killer application. Computation requirements were not there. Traditional Von Neumann computer architectures were not efficient for pattern recognition. Each one of these things have gone away now.
There is a requirement for a different architecture, with instruction and data flowing in. Computer architectures are a long way from human brain pattern recognition and power dissipation. A large number of computer cycles are required to perform the same level of pattern recognition as the human brain.
New chips have a large share of neural networks. Neural networks are a fundamental building block for AI-related machine learning. Over 300 million smartphones had some form of neural networks in 2017. In 2018, we had 800,000 AI accelerators shipped to data centres. Every day, nearly 700 million people use some form of digital assistant, like Siri, Amazon Echo and others.
Unique opportunity for India
India has a unique opportunity here. The Indian industrial companies are among the early adopters of AI. Early adopters have been defined by the Boston Consulting Group (BCG) as businesses that have fully implemented more than one AI use case in multiple industrial operation areas. India ranks third, at 19 percent, after leaders, USA, at 25 percent, and China, 23 percent.
The number of fabless start-ups in India has been growing. Its driven by applications that are non-traditional computing, that involve some form of alternative processing involving AI. India has the Semiconductor Fabless Accelerator Lab (SFAL), launched by IESA in Dec. 2018 and the Fabless Chip Design Incubator (FabCI), launched in 2008 by IIT Hyderabad.
SFAL is a Government of Karnataka initiative, which plans to allocate 20 start-ups over the next three years and 50 start-ups over the next five years. FabCI is funded by the Ministry of Electronics and IT, with Mentor and others as technology partners. It has a goal to incubate at least 50 Make in India chip design companies.
Across India, the IITs are supporting AI education. IIT Hyderabad offers a Master’s degree in AI. It is adding a Bachelor’s program in AI during 2019-20. IIT Kharagpur has a six-month AI certificate course. IIT Madras has the Robert Bosch Centre for Data Science and AI. India has a strong AI skillset. LinkedIn data shows the growth of worldwide AI skills. India ranks third, behind the USA and China among countries with highest penetration of AI skills.
Actual AI chips being developed
In 2018, there were 24 fabless AI firms that were funded by VCs. The step up since, has been enormous, and it is continuing into 2019. Now, what are these companies doing? For example, AIMotive Gmbh is doing ADAS and acceleration chip, and Beijing Intengine is developing AI chips for end-point devices across multiple industries. Hailo Technologies is making AI chips for data centre edge devices, and NextVPU is developing computer vision chip for robotics and unmanned vehicles.
The largest percentage of these chips are designed for pattern recognition. How do you verify that the chip is correctly identifying the patterns that it is supposed to identify. There are over 150 companies worldwide, developing autonomous cars and doing some sort of pattern recognition. Akio Toyota, CEO of Toyota, said that you have to drive over 8 billion miles to do the requisite testing. Therefore, physical testing must yield to virtual verification. It has to be stimulated by computers.
Now, we are semiconductor guys. We hook it to an emulator. The data goes in, and you process the image, and do things. However, that doesn’t do the whole thing. Verification platforms are in use today by folks developing driverless cars. Both, electrical and mechanical must be verified virtually.
Next, how will AI understand the multitude of gesture exchanges? It determines the interactions between people and autonomous vehicles. It understands the subtle hand gestures and emotions. Again, there is a need for verification.
The next largest area is the data centre accelerator. New companies are entering the world of semiconductor design. There is a real revolution going on. The third area is edge computing. Intelligence in electronics flows downhill. We need to force the intelligence downwards. That means, developing smarter chips.
Many small companies are developing AI chips. The methodology has changed. It is being done with high-level synthesis (HLS). Market drivers for HLS include reduced time to market with good QofR, they require FPGA prototype and SoC/ASIC, late changing specifications, and reduced verification and debug cost/time. Companies do that by writing the differentiating code in C++ and automatically synthesizing it in RTL. Example, Nvidia, which cut verification costs by 80 percent.
Can silicon tech support chip architecture innovations?
Benjamin Gompertz developed a mathematical model for time series in 1825. There was maximum growth at 36.8 percent of Asymptote. An example of this is mobile phone subscribers, which looks like an S curve. You can predict the evolution of these things.
How is the consumption of chips going to do in server centres? The growth rate is increasing. Over 4 percent of the chips in the world are going into servers. But, for how long will silicon transistor be the tool that we use to solve problems? The total transistors in Gompertz’s curve is heading upward. The growth rate for transistors will continue to increase till 2038. We won’t hit saturation until sometime in the 2050s. Data suggests that maybe, they will solve a lot of problems.
The run rate for VC spend for domain-specific chips has also increased to almost 5X in the past year. Almost half of the designs are AI-related. As many as 14 fabless AI chips were designed and released in 2018. Here, the Chinese funding is overwhelming, about six times that of the USA.
New fabless semiconductor start-ups are dominated by AI-related designs, such as vision, pattern recognition, data centre chips for cloud and edge computing. India has unique strengths to be a worldwide leader in AI and ML. The next generation of chip design technology is enabling new domain specific competitors. We can do a great deal more than what we are doing now.
By Aanchal Ghatak & Pradeep Chakraborty