“After a compounded annual growth rate (CAGR) of 3% for several years, the last two years have seen a 22% IC growth and R&D spending increased 9.8% in 2017 with an increase in venture capital funding of fabless semiconductor companies” the CEO noted.
Showing a graph retracing the yearly venture capital investments in fabless semiconductor startups since 2000, Rhines pointed that while 20 years ago, the average was circa $2.5 billion, it plateaued to an average of $1.7 billion between 2002 and 2007, then continued to decline to less than $382 over the 2013 to 2016 period. But the graph then suddenly hits $946 million in 2017 and $1.8 billion in 2018.
This is thanks to new entrants in the IC design game, the CEO interprets, IT companies are designing their own chips, automotive companies and it seems everybody wants to design ICs now. As an example, Tesla says it has to go to custom ICs if it wants to keep up with the performance requirements necessary for future autonomous vehicles. In doing so, it will replace general purpose GPUs and anticipates a 10X increase in performance over today’s technologies.
But one big driver for new chip designs comes from China with growing government incentives for semiconductor investments. Nearly $120 billion were invested over the last 5 years (of which $20 billion government backed), and China is about to further accelerate the formation of new startups with an additional $47 billion government backed fund, according to reports from the Wall Street journal.
“IC design enterprises in China used to be few” Rhines commented on another graph, until around 2000. The next 10 years on the graph see a solid average of about 500 IC design companies, surging to nearly 1400 today between 2016 and 2018.
“Their size is going bigger too, with 43% having from 100 to 500 employees” said Rhines, noting that by far, the designs of these startups are dominated by domain specific architectures, with AI and Machine Learning largely ahead of the pack (almost by a factor of 3 compared to the nearest domain, Crypto Currency).
Rhines who had worked on AI as early as in the 80’s was keen to highlight that time was now ripe for the adoption of AI, with abundant computing power and large data sets to tap into (all missing in the early days of AI).
In 2018 alone, 14 AI companies were formed and in Q2 2018, such companies cumulated $1.376 billion. In fact, early round fabless funding in China passed the US in 2018, literally dwarfing any effort from competition.
“The biggest thing they do, pattern recognition, next the second largest category is for data analytics” continued the CEO about the domain-specific AI/Deep Learning ASICs, noting that the likes of Amazon, Google, Facebook or Alibaba all want better data analytics to improve their services, they also want to push edge computing for pre-processing newly generated data.
“These domain-specific architectures require new tools and design methodologies. A large share of companies doing AI chips use high-level synthesis, because they can retarget the same C code to different nodes and FPGAs”, boasted Rhines, hinting that Mentor’s tools are first in line to help with these new designs. “High-level synthesis is about four times faster than traditional manual RTL” he added, quoting nVidia improving its productivity by 50% and cutting its verification costs by 80% using high-level synthesis for a 10M gates video decoder on the Tegra X1. The company even converted its VP9/H.265 converter from 8 to 10 bit colour in a matter of weeks, and then re-optimized the IP from 20nm at 500MHz to 28nm at 800MHz in just three days.
Of course Mentor sees all this IC design activity increase as a boon, but has the semiconductor industry reached maturity? Are the days of transistors as we know them numbered? The CEO asked as if worried his tools would soon cease to be relevant.
Rhines then came back with his favourite predictive formula, the Gompertz curve, which he fitted to historical data about the yearly cumulative transistor unit volumes to date. 2020 was only at the very beginning of the curve, meaning that silicon transistors where still due to grow in numbers, with a peak growth identified in 2018 only starting to slow around 2045.
“How long is our future with transistors and do we need an alternative for silicon transistors? The answer is, not for a while”, concluded the CEO.
“What replaces Moore’s law is the learning curve that refers to all the methods of costs reduction and performance improvements, such as growing circuits in the vertical dimension, 3D packaging and novel design strategies” Rhines continued.
“TSMC had a meeting two weeks ago and it had a clear roadmap for ramping up 7nm, while 5nm is considered pretty stable already. 3nm is on the way. Meanwhile, we continue to raise the level of abstraction for IC design. When I joined Mentor, all the layout were done in a manual sketch on paper, we had to calculate the width and the length of the gates, digitize the layout, make a mask, it didn’t work and we would rework it. Today, none of the people involved in design edit at transistor-level, we don’t train people to layout transistors, the computer does that”.
This level of abstraction is only going to increase hinted the CEO, unveiling that about 28 projects currently in progress at Mentor are using machine learning and AI-based techniques to improve the company’s IC Verification and Sign-off tool Calibre.
Mentor – www.mentor.com