The Design Automation Conference is in Las Vegas this year. If you are here and want my recommendations, then see my post Top 10 Reasons to Go to DAC. If you are not here, and want to know some of what you missed, then this post covers a lot of what I saw.

One piece of news announced today is that DAC will co-locate with SEMICON West in San Francisco in July 2020 and 2021. This is an initial two-year commitment announced by the sponsors of DAC, who are the IEEE Council on Electronic Design Automation (CEDA), and ACM Special Interest Group on Design Automation (SIGDA).

Automation and the Gaming Industry

Since DAC is in Vegas, it seemed appropriate to open the conference with a keynote on the gaming industry. Mark Yoseloff of University of Nevada at Las Vegas (UNLV) presented How Automation Has Changed the Gaming Industry. He has an unusual background, having worked on the 4004 at Intel, before moving into automation for slot machines, and now has a research group at the university studying various aspects of gaming.

He had a model of technological change that went from discovery, to invention, to innovation, to development. For example, semiconduction was first noticed by Volta in 1782,  rediscovered by Faraday in the 1800s. But the discovery on its own was not useful until the invention of the transistor at Bell Labs in 1947. The innovation phase was the invention of the integrated circuit at Fairchild in 1960. Then the microprocessor and other developments came along in the 1970s.

He moved on to talk about the evolution of casino games. He first got involved in the industry by starting a company to write software for slot machines. He picked that industry because there's a lot of money in slot machines, and gambling involves a nice combination of psychology and technology.

The oldest form of gambling was dice in 3000 BC. There are dice being played with in the ruins of Pompeii, and also some loaded dice, so cheating goes back just as far. Then there was no innovation until playing cards in about 800-1300 (and tiles were invented in China around the same time). In 1700s, the first wheel games like roulette were invented by Pascal, one of the fathers of probability theory.

The fourth historic step was the slot machine, invented in 1898. It was significant because it was the first form of automated gambling that didn't need a dealer. The general trend has been for more automation. If you looked at a casino 50 years ago, it was just those original four categories: dice, cards, roulette, slots.

He showed a picture of a wooden model of what was the prototype for the first card shuffling machine. It took ten years for this to become an "overnight" success. The first machine went into Bally's Las Vegas. It split the deck and then riffled the cards together. It did this six times, although Mark discovered when he modeled it that it really needed seven to be completely random. Now the automatic shuffling machines are different. They generate random numbers, read the cards, and put them in order. There's even an audit so that the pack used can be checked later if anything is suspicious.

Automatic shuffling led to new specialty table games. Before, with dealers shuffling, it was too risky to have big jackpots like 25,000 to 1. Games were limited to small odds like 2:1. But with automatic shuffling it became impossible to cheat so this opened the market for specialty games in the late 1980s, that all featured big jackpots.

Slot machines were always automated but others required dealers. Conventional wisdom was that people had to have chips, and see the cards, but now there is more automation of casino games. For example, in Macau, they have a huge room with four dealers dealing baccarat. But 400 people can wager on one or all of the hands electronically.

There is also now automated roulette with rudimentary robotics combined with automated wagering. Coming soon, fully automated craps (dice).

What's next? He thinks it will involve more automation and more intelligence.

I have two students who started out majoring in CS until they joined our program and discovered they had a knack for skill-based games. I think the up-and-coming gamblers who have grown up on iPhones, they are going to be challenged by this sort of app style gaming.

IoT Security

The other opening-day keynote was Microsoft's Galen Hunt on Securing the Billions of Devices Around Us. He says were are starting on the fifth generation of computing after mainframes, minicomputers, PCs, mobile, and now microcontroller-based devices. There are 9 billion of these deployed every year, everything from fitbits to refrigerators. IoT enables a digital feedback loop, connecting an organization to its customers, operations, products, and employees.

That's a great opportunity, but Galen's second emotion is fear. When you connect something to the internet, it creates a great deal of risk. "The internet is a cauldron of evil." He discussed the Mirai botnet that I covered in Video Cameras: No Service for You. A group of hackers took about 100,000 everyday appliances, ironically most of them security cameras, and turned them into a botnet that they used to take down the East Coast's internet for a whole day.

It is a challenge to build a secure device. Microsoft found this out when they build the 2001 Xbox 1, one of the most secure devices ever built. But it was hacked three weeks after it was shipped. It had a couple of what Galen regards as the seven properties of secure devices. 

They spent a lot of money building a custom security chip for the 2005 Xbox 360, which had 6 of the 7 properties. It was hacked three months after it shipped.

The 2013 Xbox One has all seven properties and hasn't yet been hacked.

Of course, the Xbox is a big powerful product built by a major corporation. Is it necessary for smaller devices to be secure? Well, the Mirai botnet pretty much showed that. But Galen had a Vegas story. One casino has a large aquarium, and someone bought a connected thermometer to monitor it. Not much security, and hackers got into it, and from there they scanned the whole network, and extracted the high-roller database. Through the thermometer.

So, his last comment was that we have a responsibility to make IoT devices secure.

Wally Rhines

Wally Rhines, the CEO emeritus of Mentor, presented Fundamental Shifts in the Electronic Ecosystem. As he pointed out, he was presenting in the time slot that Gary Smith would present when he was still alive. So Wally decided to see what's changed since the last time that Gary presented. One big thing is that we've had an acceleration in the growth of the semiconductor industry. For 15 years it had grown at an average of 3% and so wasn't really a growth industry. Then we suddenly had two monster growth years. This year was forecast to be 5% growth, but the latest estimates are that it will shrink by 6.1%. So perhaps it was a two-year phenomenon. The big increase was driven by memory price increases (not particularly increases in volume shipped, just higher prices). Non-memory semiconductor growth is likely to be flat or negative. But the big trends are in the right direction, with ICs capturing an increasing share of electronic systems' product value. Average in 1992-2014 was 16%, but now it's up to 22%.

Another big change is who is doing the design. System companies like Google, Facebook, Alibaba, and Tesla need to build their own proprietary ICs. Another wrinkle is that these chips are not really captured in the industry growth numbers. They show up as foundry wafers, but at a much lower value that they would show if they had been built by a fabless semiconductor company and then sold to the system company. In a sense, they only show up as wholesale price, not retail. Despite this, system companies are the fastest growing part of the market, with a CAGR of over 70% (starting from a tiny base a few years ago).

When Gary last presented, there had been a worry that there would be a lot of consolidation of the semiconductor industry. For instance, Mark Edelstone of Morgan Stanley predicted half the public semiconductor companies would be acquired, and the share of the top three suppliers would double. But since 1965, in fact, there has been deconsolidation, the share of the top 50 companies was 98.6% and now is 83.4%. Things are changing fast since the memory companies are declining faster than the top 25, so deconsolidation is starting again. This is good since a consolidating industry is mature and doesn't spend much on R&D. But semiconductor R&D is constantly 14%, with 2% consistently going to the EDA industry.

Semiconductor growth comes from new applications, PCs, mobile. The big question is whether IoT will create a new wave? It is clearly a new age for processors, since most of the investment in processor design has historically gone to make life easier for the software people, who could just wait for new faster nodes to come along. But that is over and domain-specific processors are driving a new wave of semiconductor growth, with huge processors for the data center, and tiny ones for the edge. Traditional von Neumann architectures are just not efficient for pattern recognition.

As a result, the secular decline in fabless semiconductor investment has reversed. It was $2.5B in 2000, then down down down. But it was $2.4B in 2018, and is nearly $1B year-to-date this year.

Looking at the applications that these startups are working on, the biggest is pattern recognition (face, sound etc) in green in the above chart. The biggest are SenseTime and Face++ with $1.6B and $0.6B invested in them. The second biggest segment is stuff to go in the data center to accelerate things. Then edge computing.

One question is how these startups afford the cost of IC design? Five years ago, we were saying that we would be selling to just a few big customers, the ones who could afford design. But all sorts of people are doing 7nm designs today. It is 25% of TSMC's revenue. 

Gary Smith would appreciate one of the reasons, that is a change in the level of abstraction. The system companies that are the fastest growing segment have no legacy design methodology and they are all using high-level synthesis. The differentiation is in the high-level algorithms and in the dataapth, so they start with C++ and automatically get huge reductions in time to market. They also get flexibility, making it easy to put the same design in an FPGA, in a 28nm chip, in a 7nm chip, starting from the same design representation. The verification is much faster.

Google's experience, summarized on the above slide, is that they were going 500X times as fast as with the prior RTL methodology. Three servers for two hours could do what 70 servers would take two days to do with RTL. And they caught almost all the bugs at the C level, too.

The power is a lot less, too. As Wally put it:

If you can do 500 times more simulations, you are going to find an architecture which is power efficient. At the architecture level, you can get a big impact.

Wally wonders why people didn't do this sooner. After all, high-level synthesis technology has been around for 10 or 15 years. He thinks it is probably having good formal tools so that you can trust that the RTL really does match the untimed C++ you started with.

Wally's industry summary to wrap up:

Last four years since Gary died have been a remarkable period with all the system companies joining in, and despite the fact we will have negative growth this year, the investment is continuing, which is good for all of us.

Analog Lunch

Since Cadence announced the Spectre X simulator this morning (for details on that, see my post Spectre X: Same Accuracy, New Speed), we moved the order of our lunches around and had analog/mixed-signal today. I'll write about that in a separate post, but to whet your appetite, the topic was Closing Analog and Mixed-Signal Verification in 5G, HPC, and Automotive. Professor Georges Gielen of KU Leuven was the moderator, with panelists YY Chen of Mediatek, Atul Bhargava of ST Microelectronics, Roopashree HM of Texas Instruments, and Cadence's circuit simulation R&D manager Vinod Kariat.

Cooley Troublemakers

The traditional John Cooley Troublemaker Panel was also today. I'll cover that in a separate post since this one is already so long. The panel consisted of Joe Costello of Montana (hardware accelerated SystemVerilog), Naveed Sherwani of SiFive (RISC-V processors), Cadence's president Anirudh Devgan, Joe Sawicki of Mentor, Mo Faisal of Movellus (Analog IP), and Raik Brinkmann of OneSpin (formal verification). Of course, John was the moderator, although as Joe Sawicki pointed out, "it makes things interesting when the troll on the panel is the moderator."

Sign up to get the weekly Breakfast Bytes email: