Monday At DAC

The 54^th DAC got started today in a very steamy Austin. While we may be a maturing industry, there is certainly no indications that the people within the industry have given up or intend to take it easy. The event really got started late Sunday when Laurie Balch, chief analyst for Gary Smith EDA, delivered her message. She said that the focus is becoming the verticals. “This change in focus is in part caused by the slowdown in migration to new technologies and will create more demand for tall thin engineers. This in turn may require a refocusing of tools, which today are targeted towards experts.” She also said that we need to look at and learn from adjacent industries, such as mechanical, that have already gone through the slowdown that EDA is facing.

Joe Costello keynote
Monday started by trying to create some of the dynamism from the past and Joe Costello, the force that created Cadence and its first CEO, now leads an IoT company. “I believed that the IoT will be bigger than anything we have ever seen before. The Internet caused a degree of consolidation of the industry into a few devices and this led to a decreasing number of successful companies producing them. The IoT will be the opposite. Things will become diversified and each will require different technologies, different sensors. We will see proliferation of design.”

Costello also said that many failed IoT companies are focusing on the wrong thing. “Everything will be a feature and not a product. IoT is about intelligence. Early products gave you control, but this is not the IoT.”

He also provided a few failed examples, including an automated dispenser of dog food. While there were a few chuckles at this reference to the past, I think the industry has said goodbye to many of the people who would have understood this joke.

He then went into more detail about what his company is doing and the approach they are taking. It was fully recorded so I am sure that DAC will make the entire talk available in the near future.

Machine Learning and EDA
The next session attended was a panel about machine learning and EDA. The abstract posited that fast growing semiconductor market segments, including high-performance computing, automotive, mobile, and internet of things (IoT), are driving the need for electronic design automation (EDA) software tools to advance even more quickly in addressing increased complexity. It asked the question about how EDA is using this technology and where it is helping.

Ting Ku from NVIDIA set the stage by looking at the degrees of automation. “In the broadest scope we have rule based automation. Within this is machine learning and deep learning. The difference is that rules are deterministic. Learning is statistical and deep learning has no predefined features. Tools learn by iteration from its previous decisions. It does not need to be neural but that is the area of most interest. EDA needs to stop just providing data. We need well informed decisions.”

“Most of the potential is in places such as variation aware design, memory and power estimation,” said Eric Hall from E3 Data Science. “It reduces uncertainty and the padding necessary. With accurate estimation you can create an efficient frontier for things such as memory size. Most of the models used so far have problems, such as linear regression, that is not good at handling discontinuities.”

Sorin Dobre from Qualcomm identified opportunities for optimization in the design flow, IT resource allocation, IP characterization and data management. “The flows are automated but rely on engineering experience. When you have a new design team you cannot guarantee the quality of the design. We need better ways to extract knowledge from one design to use in the next one.”

Jeff Dyck from Solido Design Automation, provided a glimpse into their experiences with machine learning. “We can smell machine learning problems, but we can’t just take a course and use that to solve problems. We certainly have massive data and most of that is streaming in real time. The data is complicated and has multiple dimensions. We cannot base designs on guesses, we need higher levels of confidence.” He also saw additional challenges for EDA. “Most EDA algorithms do not port to parallelized architectures very well. Incorrect answers can pollute your learning. You need to be able to debug real time data. Engineering requires the right answers.”

Synopsys lunch
Today’s lunch was provided by Synopsys in exchange for listening to them talk about accelerating robust AMS design at advanced nodes. The problem is simple, noted the speaker. The number of elements is increasing rapidly, and at the same time the number of coupled capacitances is growing even faster as we go down to smaller nodes. If you thought analog simulation was bad in the past, then it will get a whole bunch worse unless you start to add more intelligence to solving the problem. “We need to become a lot more careful in terms of what simulations are done, how they are done, and how much data is created,” said Tom Mahatdejkul from ARM. “These were not important in the past but today is become so difficult that intelligence has to be added into the process.”

The IP Paradox
Another panel got off to a rocky start, but posed an important question. If the semiconductor industry is not growing and companies are consolidating, how come the IP industry is still growing? Many aspects of this are well known and a lot of old ground was covered. It got more interesting when asked if consolidation was impacting the IP market. “It has been a tailwind for the industry,” claimed John Koeter from Synopsys. “When they are consolidating they are looking at what differentiates them. With the rapid evolution of things such as USB, these are getting complex but they are difficult. Fabs are also developing new process technologies and the cost of supporting these is high.”

But Sanjive Agarwala from TI, a user of IP, was having none of that. “There is a feeling that IP is growing, but it will go the way of EDA very quickly. Number of nodes will slow down, there is consolidation, there is standardization. It will saturate in a few years and it will be tough to operate there. IP companies cannot raise prices.”

Elias Lozano from Open Silicon was somewhat in the middle. “The cost of IP development is huge but if they can amortize it across enough users they will continue to do well.”

Agarwala struck back. “It is not just technical, it is business. The business model is changing.”

Koeter was still having none of it. “We study the market and 60-70% of the IP is outsourced. When I look at IP, I think it is potentially the same size as the EDA market. EDA is fully outsourced, but IP is not there yet which means there is growth available. IP does not stop with a title, it does not become commoditized because they are always asking us to provide more value around the IP, such as software and optimization. When you can create value, you create revenue.”

Acquisitions in the industry
And for anyone wondering how Mentor would be presenting themselves at their first DAC after being acquired by Siemens, I can tell you that they are still here in force. I caught up with Wally Rhines, who looked a lot more relaxed than he has for a long time, and asked if he had any thoughts about how Cadence and Synopsys now viewed Mentor as a competitor. He answered that he believes both of them will have to take them a lot more seriously because Mentor now has much deeper pockets standing behind them than either of them.

Mentor also wined, dined and entertained us for the evening. Rhines reviewed his previous year’s predictions related to the probabilities of mergers. Last year he had talked about the most likely acquirer was a company with the first letter being M and the most likely acquire being a company starting with the letter A. He admitted that their analysis was flawed and that they did not pay enough attention to the second place acquirer being that of S and their probability of acquiring an M company. If he had have been, he would have been a lot more suspicious about Synopsys. But the rest is history.

Further Reading
Tuesday At DAC
Why did Siemens acquire Mentor and who will be next? Portable stimulus contemplates a tough adoption challenge.
Wednesday At DAC
What will be the next killer app for the IoT, establishing C++ as the signoff model, new architectural challenges and tackling verification challenges.