New Data Management Challenges

An explosion in semiconductor design and manufacturing data, and the expanding use of chips in safety-critical and mission-critical applications, is prompting chipmakers to collect and manage that data more effectively in order to improve overall performance and reliability.

This collection of data reveals a number of challenges with no simple solutions. Data may be siloed and inconsistent, which means not all data is usable across the supply chain by all the relevant players. No clear rules exist for how long that data should be stored, who should have access to it and for how long, or how that data can be looped back into the flow to improve quality and tradeoffs involving power, performance, and area. Also unresolved is how to quantify and qualify data, how to best learn from it, or when and how to make it available.

Various commercial tools do exist to help manage design and IP data. But the quantity of data that needs to be managed, tracked through IP reuse, and leveraged at different points in time by different groups is a mind-boggling challenge.

“More and more data has be managed, or has to be found,” said Simon Rance, vice president of marketing at ClioSoft. “Depending on the size of the company, that data could be on one continent or different continents. You may not know where all the design data is, so there’s a need for both design data management systems and for improved collaboration. How do you re-use those designs? Do you know the quality of that data so you can apply it to a future design? That’s where we’re seeing that transition now in the data management space.”

Data management needs to happen on multiple levels, starting with the design and all the way through manufacturing and into the field. But as connections are made across data, which may not be in one place, management can become very complicated.

“When you’re in the small data world, you run your wafers, you get the results back for those wafers,” said. Rob Conant, vice president, software and ecosystem at Infineon Technologies. “You run the tests, you aspire to a certain level of reliability or performance, for example. Then you start shipping hundreds of millions or billions of those things out into the world. But in the semiconductor industry, historically — as well as pretty much every industry under the sun — your ability to monitor those things disappears. So you hit print, and boom, it’s gone. It’s like selling [an operating system] in 1995. We shipped a bunch of disks, everything seemed good, we got some support calls.”

Much has changed since then. “The whole software world has radically changed so that people are watching exactly what you’re doing,” Conant said. “All of a sudden, you’re measuring billions of people using these products, and it radically changes the way those software companies design their software because they can actually see what’s important and what’s not important, where people are running into problems, what’s working, what’s popular. It totally changes the software design methodology.”

Much of this is now being adopted by the semiconductor ecosystem. “We’re still going to be shipping hardware, but we’re measuring, for instance, this die went to that customer and then eventually did that. And we’re doing that hundreds of millions of times. Is that data going to feed back into our manufacturing process? Are we going to be able to say, ‘That wafer, at that facility, on that date had a lower reliability than this wafer, at this facility, on this date?’ Probably. That’s the direction things are going.”

Kam Kittrell, product manager for the digital sign off group at Cadence, noted that in order to do that type of telemetry, collaboration is needed across the industry. “For example, when performing an IR drop test on the design — which is a test of the veracity of the power — if I’ve got a few things switching, will I spike my power grid to the point to where it could dangerously affect the function of the PHYs? But because you can put any device into a critical power drop, if you’re switching it up in the same area, what you statistically check may not be what was realistically happening. There are ways to add voltage meters and drop them in parts of the grid, so you can go back later and scan it to determine the low-water mark of the voltage. This provides feedback that can influence design decisions on how to avert this in the future.”

What is seen on the tester may be significantly different from what’s on the board. “Once the product is live, users start to make requests for machine learning as they keep trying to optimize the design,” said Jerome Toublanc, business development executive at Ansys. “With simulation data, for example, when it comes to managing that data, we focus on what matters most, and this is where a simulation data management process comes into play. In the past, users were keen just to collect data and use that data during the design process until tape-out. Now, users want to include more data coming from the lab, and account for data from the field. Now they want to include data from their product, be it 5 years or 10 years in the product, because they will learn even more about the quality of the product. This is where we see an increasing amount of data, compared to where we were even 5 or 6 years ago. Then, no one was asking us to include the data management data from the product once it was in production. That’s a completely new way of managing data and using it.”

Toublanc noted that a number of IP companies are providing IP for in-field data gathering. “That speaks to part of the product lifecycle. Once you’ve gathered that data, you can see how it performs over time and then correlate that back to the assumptions when you did the design. But then you just have to deal with data growth, because you used to get a lot of data after design. Imagine collecting years of in-use field data after that.”

Thinking differently about data
In effect, the semiconductor industry is becoming an IoT industry by itself in the way the chips are deployed. “That’s a pretty radical idea that will have a huge impact on this industry over decades, in the same way that it has on industry after industry,” said Infineon’s Conant.

But how does the system architect and design team decide what data to keep, and for how long? And what is done with the data before releasing it? The answer depends on who wants that data, the analytics, and what they want to do with it.

“For the design team, they may not want to look at it right now,” said ClioSoft’s Rance. “It may be that the architect wants to look at it to improve the next architecture of the design from a performance and low-power perspective. But there’s also the other aspect of what is not working as intended. How can the software team potentially do an over-air software update to compensate for the hardware in order to take action immediately?”

This is beginning to happen today. “As part of deploying WiFi chips into the field, we started aggregating and collecting data specifically for battery operated products and what the energy consumption was of these devices in the field,” said Conant. “In one case, we found that there was an average, but 20% of the devices consumed 8 times the average. That means if you’re supposed to have a 16-month battery life, that 20% would have a two-month battery life. So you would have a bunch of angry customers. We started analyzing that data, correlating it, figuring out what it correlated with, and eventually figured out what was happening. We provided a software update to the companies that operate these products in the field. They did an over-the-air update and tripled the battery life for those outliers after the things had already shipped. That is what an IoT company does today. That’s what they do all the time to improve their products in the field, and I believe that the semiconductor industry is going to do more and more of that. In our case, we just brought it back to the software that runs on the chip. If you can bring that back into the chip design, the benefit can be even higher.”

Where to start
When to implement this kind of data sharing depends upon the market, the end product, and the various use cases.

“Most companies they have their own methods of data management,” said ClioSoft’s Rance. “It’s the same with verification teams. How do they collaborate together? How do they leverage the data? Who does what with it? And how do they come back and say what they’ve learned, what they need to improve? That’s something that needs to happen. We’re not seeing it so much yet. We see executives at big companies acknowledge it’s a problem. They know they’ve got to come up with a solution and a methodology around it. That’s where we see it going.”

In the past, built-in self-test (BiST) approaches were used in chips and in industries such as automotive, but without any consistency. “Building the system with hardware diagnostics in mind to get meaningful feedback to the hardware designer early on, so they can analyze what may be the problem and how to prevent this on the next generation and so forth — that’s the goal,” said Kittrell. “But the goal is always shifting, because the next generation probably will be a different node with different sets of problems, and so forth. They’ve got to be able to at least take those down before they cause a failure. You don’t want to be holding a customer’s payload when it goes down and is causing disruptions.”

Applying data earlier is the goal. “Users now want to start to collect data from day zero of a new project,” Toublanc said. “At the beginning, the data is very traditional, very basic. But what people like to do is to track those metrics and make sure that happens all the way through the design lifecycle. Everyone will do that in a different manner, depending on the project. This is why the solution now has to be as open as possible. Everyone will not track the same metric the same way, or even track the same metric. So our solutions have to be fully open, no matter which tool they use for design, to make sure the customer can get whatever they need and track it the way they want. This is changing the EDA solution. If you are not open, data management will be a big problem.”

This also becomes essential for digital twins, an old concept that is gaining ground as more electronics are deployed in safety- and mission-critical applications.

“Car companies don’t build cars to crash test to do studies,” said Michael Munsey, senior director of technology solutions sales at Siemens Digital Industries Software. “They build one or two for government sign-off on a crash test, but they do everything virtually. So systems companies have a good idea of what data they need to capture and keep, because they’ve been thinking about this for certain reasons. On the semiconductor side, we have the issue of generating more data than we can deal with, and this is something we can actually learn from. Some of these systems companies know the right data to keep because they have a good idea of the data they need for analysis and crash tests and things like that.”

Put simply, data collection, analysis, and retention needs to be more granular, and it needs to happen in more places within the ecosystem. “We provide a ton of components into electric vehicles, as an example,” said Conant. “Who is best in the world to model the remaining lifetime of those components? We are. But that requires the provider of the piece of equipment to also provide that digital twin in order to have those capabilities, as well as the data feed that informs the digital twin so it’s accurate on a device-by-device basis. This requires a data partnership, which the industry is not quite comfortable doing.”

This is especially true in industries like aerospace, where developers don’t want to share sensitive data. But in general, data sharing is improving.

“Previously, EDA would work with the semiconductor provider, they’d make off-the-shelf parts that would go into who knows what, wherever they could find a customer,” said Cadence’s Kittrell. “It was up to the customer to put their system together and figure out the reliability on it. Now that lots of silicon is going to very specific usage, they’re driving high volumes for each one, and it’s very important. So the systems company is pushing down how this whole process must work, looking at the data center, for instance, to be able to detect the chance of a failure. That influences the entire design process and IP selection.”

Still, not all of this data is produced in a consistent format. data needs to be interoperable. “There’s got to be some standards of the important data, and it has to be customer-driven because that’s what makes things actually move,” said Siemens’ Munsey. “There also needs to be work on a common data model with results from analysis tools, which will start allowing things to work together. There must be a way to manage and keep track of things a lot easier. That’s the next step of the evolution. You could still own the data, but at least create it in a way that can be analyzed by other people as well as the customer.”

After that, the data needs to be usable up and down the design-through-manufacturing flow.

“We’re tackling the point tool optimizations now,” Munsey said. “Then, when you have the data model, you can collect the data all the way across the board and have the place-and-route tool say, ‘If we did this in synthesis, we could have done something better here.’ That way the tools feed information back upstream, so the next time through you’re actually feeding new constraints to the synthesis tool to allow for even better results to place-and-route. You keep feeding this farther and farther up the chain, and that’s the whole idea of ‘shift left’ automation, which is back to the question of what data do you want to keep? You keep as much of it as possible because you might not know what data you need to analyze until you start realizing these other optimizations. And you could then consider that maybe the place-and-route tool is actually looking at the data that was created by the synthesis tool.”

Conclusion
All signs point to sweeping changes in the semiconductor design ecosystem, from concepts to tools and methodologies, in order to leverage and learn from data.

“If we look to the example of GE and Rolls Royce providing engines, they actually moved from providing that as a piece of hardware to providing it as a service, because that inherently changes the business model to not shipping a component, but wrapping in all the data and all the maintenance around that engine,” said Conant. “They realize there’s so much economic benefit to bundling those things. The semiconductor industry hasn’t done that. I’ve never heard of a semiconductor company providing a semiconductor as a service. But it’s happening in a bunch of other industries that are higher up the stack and it begs the question, ‘How does the semiconductor industry move in that direction?'”