Semiconductor Engineering spoke with Joseph Sawicki, Executive Vice President of IC EDA at Siemens Digital Industries Software. John Kibarian, President and CEO of PDF Solutions. John Lee, general manager and vice president of the Ansys Semiconductor Business Unit. Niels Faché, Vice President and General Manager, PathWave Software Solutions, Keysight Dean Draco, President and CEO, IC Manage. Simon Segars is the former CEO of Arm and a board member of Vodafone. and Prakash Narain, President and CEO of Real Intent. Below is an excerpt of a conversation that took place in front of a live audience at the ESD Alliance Annual Meeting.
SE: What are the biggest problems facing the chip design industry? Where are the biggest opportunities?
Lee: Multi-die heterogeneous 3D-IC systems present both the greatest opportunity and the greatest challenge. There are also big challenges regarding China, especially his EDA. There are many start-ups out there, and our ability to sell to China is extremely challenging.
Seagers: The current recession is a bit of a challenge, but there are some new technologies on the rise that offer opportunities to continue investing through them. It creates a playground for building things. But doing so will require much more complex tools, flows and methodologies than what exists today, and will require a lot of research and development. Therefore, I would like to ask, “Don’t skimp on research and development because it will be necessary in the future.”
Draco: The biggest event before us is the impact of AI on our industry and the industries we serve. AI in the automotive industry and AI in the video surveillance industry will drive enormous amounts of silicon consumption. And AI in the EDA industry will be huge. We will develop all these chips, we will be the AI consumers that manufacture the chips, and we will use the tools to manufacture the AI chips. The appetite for AI will drive the move to the cloud faster than the EDA industry. Our customers want to move to the cloud. Because customers are hungry for GPUs (and the systems that deploy them in large numbers) to run it all. Basically compute to AI root or AI design chip. And that AI needs a lot of data. The chip designs we are designing today are terabytes of data. With the introduction of AI components, the data will be terabytes. That creates this crazy hard problem, and we’re going to work on it and solve it. But when you start doing that, you end up with a lot of data on-premises and in the cloud, some of which you need here and some there. You could do it cheaper with Google, but the data is still on Amazon, etc. So you’re going to have a big problem of data movement. There is also an opportunity to create tools that can project data to display it where it is not and where it can be conveniently and quickly displayed. AI is both an opportunity and a challenge, and data management to make that AI a reality will be a big challenge for us to overcome.
Sawicky: We are no longer suffering from a lack of challenge. After spending decades in ignorance of what semiconductors are, our government suddenly realized it was the most interesting subject they could imagine tackling. It’s a challenge for us. But we’re in this really great place where we’ll be able to monetize what we’re doing and really change the world. It’s an incredibly exciting time, but it’s also a scary time.
Fasche: One of the challenges we recognize is that designing, testing, and building systems such as telecommunications networks, cloud infrastructure, and electric vehicles is becoming increasingly complex due to performance requirements, new types of functionality, and new technologies. It means that As a result, product development teams and the entire supply chain of any system have placed more emphasis on virtual prototyping. They are really shifting the product development lifecycle to the left, allowing the complexity of systems, subsystems and components to be addressed upfront and within the product. It can also reduce time to market, improve productivity, and reduce costs and risks. Achieving this transformation requires digitizing engineering workflows, from requirements to the process of manufacturing a product to meeting specifications. This digital transformation presents a huge opportunity for our industry. It requires an open ecosystem with connected design, simulation and test tools and intelligent workflow automation.
Narain: The design workhorse is the most widely deployed verification and simulation, followed by formal and static approval, and this is where we stand. The biggest opportunity for static signoff is a shift left in verification. That is design validation at the earliest possible stage. If so defined as a design step, the designer should be involved. So there is a lot of time pressure to be successful with these applications so that the best possible user experience can be created. There are enormous opportunities to shift left in design flows, and these applications need to be very timely and very efficient. The challenge is economic innovation. We continue to invest in technology and innovation to design at the speed you need. We also continue to expand the reach of static sign-off through product and technology advancements.
Kyvarian: Manufacturing will benefit from the next node of Moore’s Law, Dennard Scaling, plus focus all manufacturing on a few very good manufacturers, all on the front end of the wafer fab have enjoyed the benefits of very controlled way. Dennard scaling isn’t happening anymore now, and hasn’t been for the last decade, but as Moore’s Law slows down, advanced packages should have more heterogeneous systems. The manufacturing challenges arise because the value is not just in the wafer fab. Assembly has become a very difficult process and test points are becoming more complex. Therefore, the production flow has an opportunity to improve yield. Plus, we’re using a variety of silicon and we’re starting to use more on-chip silicon photonics and other technologies to give us higher yields and more watts per dollar regardless of what we do. Continue to improve per performance. Dennard Scaling or Moore’s Law. This poses a big challenge for the manufacturing floor. In addition, geopolitical reasons are currently fragmenting and moving supply chains. I cringe every time I see Morris Chan talk about how Americans can’t manufacture and how the CHIPS Act is a $52 billion waste. Looking at historical data, he is correct. But when EDA officials look at the problem, it’s similar to how Google looks at the advertising industry. They didn’t just have a ton of bots peddling ads. they took a different approach. And there’s a much different approach that we can put into manufacturing. That means software and his EDA will be the tipping point.
SE: What do you think the future holds for AI in design? Who will use it? Will we get better results from this? And what are the challenges to be tackled in the future?
Seagers: AI can help significantly improve productivity. Probably everyone has played with his ChatGPT and typed something and was very surprised by the output. You can ask it to work with all sorts of different languages and what it produces is pretty good. Interestingly, GitHub has introduced this code popup, which is an integration with Microsoft, and the demo looks pretty cool. You’ll be more productive by putting more code together and eliminating errors. How many of us have banged our heads against the screen looking for the missing closing parenthesis? In other words, everything doesn’t work. With lots of auto-generated code, all such things may be a thing of the past. In the short term, there are a lot of automations that can help you get to the point where you really need to use your brain power. And in the long run, it will be even better. On the other hand, just writing test cases and understanding what flaws are in the test cases also helps with validation. This is what we started a long time ago. Using machine learning to understand what is good and what is bad from a validation perspective can save a lot of time, reduce computation cycles, and get you to the finish line faster. It would be pretty revolutionary.
Draco: 10, 20, 30 years ago everyone used the term expert system. I can no longer use it. But we convene a large team to do the validation. That team consists of her 5, 10 or 100 engineers who have experience in verification and 5, 10 or 100 engineers who are less experienced or less talented. And we will train you how to write test cases. This may not be necessary if you use tools similar to ChatGPT validation provided by Cadence, Synopsys, or the company that created their expert system. You may be able to generate a large number of test cases, but 10% of them may be a little off and you need to fix them. But increased productivity makes a big difference. When they first envisioned AI, people thought it would threaten and take jobs, and blue-collar workers immediately came to mind. That’s completely wrong. AI threatens white-collar jobs. I don’t think we need an attorney anymore. Just stating the path to ChatGPT produces a pretty good pattern. No need to spend $20,000 on that law firm. Harnessing AI in many ways presents an opportunity to improve the productivity of white-collar jobs.
Sawicky: Over the next eight years, Moore’s Law is slowing down, but consolidation is still going fast. Considering the breadth of solutions we need to put together, and considering that every 2-3 years he expects to improve 2x, that stack is even bigger. So there’s a huge system-level context where the AI chip isn’t just doing something locally. Whole systems come together, like factory automation or driving a car, and they all work together. The scope of verification is vast. Given the state of education in the United States and the number of engineers that will graduate in the next eight years with existing resources, can you imagine we’re going to get there? Thank God, we are in a place where we can take advantage of the opportunities in front of us. From a chip design perspective, what would happen if generative AI entered the design space? How would it help? How could it be revolutionary? How do you discover what’s been done? What’s the interesting connectivity? What’s put together in a really compelling way? there is. How can we innovate based on the baseline we need to do with our tools? It would be a great undertaking in terms of how we take advantage of these opportunities.
SE: We are dealing with much more data than before. Is that data okay? And how do we know it’s good?
Kyvarian: One of the big opportunities for AI and ML is improving the data it works with. We are seeing it happening more and more. Tools help ensure the consistency of data coming from the machines that build software, the test machines. It will be deployed and people will use it even more in the next decade. It’s not that far.
Lee: There is a very interesting set of activities that not only ingest data, but can also extend that data. Therefore, the digital twin of the manufacturing equipment will contain everything that was sent, including modeling errors. But combining it with data from manufacturing is a great way to make practical self-correction.
Draco: But AI data will be very problematic in practice. Because each design company wants to control it and not put it elsewhere. They are cautious, paranoid and beyond control. Therefore, there is a significant challenge for the EDA industry to build the AI performance tools that the design industry requires without access to data or acquisition of data to train the tools and neural networks.
