More Data, More Processing, More Chips

Simon Segars, CEO of Arm, sat down with Semiconductor Engineering to talk about the impact of heterogeneous computing and new packaging approaches on IP, the need for more security, and how 5G and the edge will impact compute architectures and the chip industry.

SE: There are a whole bunch of new markets opening up. How does Arm plan to tackle those?

Segars: Luckily for us, we can design products that get used in lots of markets. Over the past 29 years, what we’ve done is get a good enough understanding of the end markets so our products can be used in lots of markets. Having said that, we have become increasingly vertically focused because there is an increasing amount of specialization in end markets. You can build products and hope they get used everywhere, like in mobile. But as we look at enterprise or IoT or automotive, you need partners, third-party IP and ecosystems to create solutions that are specialized for an end market. There is more performance, more efficiency, better power. It’s moving a long way from one size fits all. It actually was never that, but you need a lot more specialization today.

SE: Arm has been expanding its portfolio for some time. Will that continue, or will the focus be on putting the pieces together as modules?

Segars: It’s not modular. The way we look at our CPU roadmap is there is a core processing engine, which requires some specialization for different markets. So if you’re going into the enterprise space, you want error detection and correction, and there may be a set of features you want to add on top of this core. If it’s automotive, there might be redundancy features you want to add. So our core engineering is effectively one big team producing variants for different markets off a central piece of work, which is essentially the core of the core.

SE: Will your processor cores remain soft, or will you harden some of them?

Segars: They will be predominantly soft, but with each generation of technology the role that our physical IP team plays becomes more important. Ultimately, our licensees build something on silicon, and it’s the performance of that chip that matters at the end of the day. Having a deep understanding of process technology and where that’s going, and how the changes there affect the processor or the processing unit, really matter.

SE: How does automotive change things, particularly with some of the thermal requirements?

Segars: We’ve been involved in automotive since about 2005. Every six months or so, the number of corners you need to characterize your library to goes up. In the past when we delivered a library, it was a fairly simple set of data. That has exploded. We now need different views and different corners. It’s unbelievably complex.

SE: There is a move toward chiplets as a way of speeding up time to market. How does that change what Arm does?

Segars: We’re starting to look at that, because 5nm is an economics problem. Creating variants on a theme is really, really expensive. If you can create a compute subsystem on a chiplet and integrate it with other dies — and that chiplet could be on advanced digital node or it could be analog — then you get some economies of scale out of designing that subsystem. The question for us is whether chiplets represent the future of IP versus soft IP.

SE: Designs are also getting more heterogeneous, as well. Is that an opportunity or a threat?

Segars: That’s a somewhat separate subject from chiplets. Going through the act of building a chip involves fixing a lot of things down and understanding the system performance of the end product that you’re trying to build. Once you build it, it is fixed. There is only so much flexibility you can have if you go down the chiplet route.

SE: And you really have to characterize all of the IP in that device, right?

Segars: Yes. We understand characterizing hard IP. We build a lot of test chips and we run software and do ESD testing, for example. So it’s a space we understand, but it’s slightly different from that chiplet model. There is no IP company shipping chiplets in volume yet. On the heterogeneous side, there is a model of a chip composed of multiple chiplets. You can have a CPU cluster, a GPU cluster, and you can communicate between them over PCI, for example. So on these advanced nodes it perhaps increases the flexibility for mix and match, as opposed to building one huge die that’s going to cost a fortune. But there are all these kinds of interconnects and delays from ‘system one’ to ‘system two’.

SE: Yes, that’s still under development.

Segars: But there is still possibility that chiplets will form a pretty significant part of chips in the future.

SE: What happens on the heterogeneous side?

Segars: Chips have been heterogeneous for a long time. Arm cores were put into heterogeneous chips in the original cell phones.

SE: But there were not as many processing elements as in today’s AI chips, right?

Segars: When you talk about AI chips, there is such a huge spectrum. Some of them are very similar to those original modem chips. There’s a CPU and an accelerator. You can put that together. To have that in a microcontroller, where you have a really scaled down accelerator next to a Cortex-M, or at the other end of the spectrum where you have a chip like Cerebras has introduced, which is basically a wafer. You’re trying to exploit massive parallelism. So there is a spectrum of multiply accumulates doing vector math at one end, or a CPU, which can be big or small, with some help doing vector math because it’s a subset of all the things that processor does. Or it could be a subset of the application that needs acceleration some of the time, so you spend a bit of silicon to improve that.

SE: So what you’re doing is prioritizing different operations?

Segars: Yes, and when we look across all of that spectrum, there’s a massive market of AI at the edge, at the end point, through the network, and machine learning. Then there is AI acceleration in the cloud, which is potentially interesting for us. But that mass market outside of the cloud is traditionally a more obvious space for us.

SE: How do you see the edge materializing? Will it be the endpoint, edge and cloud, or will it be sliced up even more?

Segars: People use the term ‘edge’ sometimes to describe the endpoint, and sometimes to describe where the endpoint gets onto the network. For us, we’re trying to be consistent in defining the edge as the edge of the network, and the endpoint as what is connected to that.

SE: So it could be a sensor, right?

Segars: It could be a sensor or it could be a car. There’s a point where you’re on-boarding to the network, where a lot of data points are going to come together. It’s also where data that is being saved up might get stored. But the edge node becomes a lot more sophisticated than it has been in the past. It’s not just a hopping-on point to get to the cloud, where all the interesting stuff happens. It is doing some of the work and storing some of the information. That becomes a large compute environment where we see a lot of opportunity.

SE: How does the edge differ from the IoT concept?

Segars: IoT a broad category. With IoT, you have lightbulbs in the home transmitting data over a hub. There isn’t a lot of data processing going on in the lightbulb. But you also can imagine streetlights talking to the edge of the network, and if an accident occurs, they can reconfigure things. That can happen much more locally than shipping it off to the cloud, where there is latency involved.

SE: What you’re talking about is massive amounts of processing everywhere.

Segars: That’s our business. We’ve shipped more than 150 billion chips over the years. All of this will drive more intelligence in devices, and that’s good for Arm.

SE: Do you have to develop new chips and cores as a result of this, or is it just a different way of putting them together?

Segars: There are definitely new opportunities emerging. The work we’ve done with ML is a whole new category. It’s emerging quickly. And if you take something like autonomous vehicles, or the way the automobile industry is developing in general, that’s getting much more sophisticated. It requires different features to provide safety-critical assurance. That brings another lens to how these design practices are changing, and how the skills of engineers need to adapt constantly.

SE: There are a lot of moving pieces in all of this.

Segars: The network is core, and 5G is an evolution of that. It will allow all different types of things to be connected in a very different way. The rollout of 5G is really important. That creates opportunities for other products, which will evolve on a different timeline. Self-driving cars will evolve slowly. I’m confident the industry will get there, but it’s not moving from nothing to self-driving cars. There’s a whole evolution that will happen. 5G enables that. It enables mobile broadband. It enables the connection of IoT devices directly, instead of going through a WiFi hub. It opens up new things, which will evolve on different timelines.

SE: Is a 5G device going to be a phone, or will it be a building where you don’t need so many small cells?

Segars: It’s going to be in a phone, for sure.

SE: But there is sub-6GHz, and there is millimeter wave, and both of those have radically different properties and restrictions.

Segars: Yes, and in your house you probably already have a broadband router, which relies on a physical connection to the broadband. You can suck in a 5G cellular signal and it will turn out WiFi on the other end. That gives you flexibility, because you don’t have to worry about where the socket is on a wall.

SE: But will we use millimeter wave technology the same way as sub-6GHz technology?

Segars: 5G is very different on a number of different fronts. It will make your phone faster, but that’s only the tip of the iceberg for what 5G will enable.

SE: Does this mean more processing and more chips?

Segars: As a market, it will consume a lot of silicon. That’s good for the semiconductor industry, and it’s good for us. Devices will require more intelligence than in the past. There is greater volume with the number of small cells that will be required. That is a good driver for the networking industry. And just deploying it is different.

SE: Let’s swap gears. How much has Arm been affected by all of the geopolitical upheaval, such as the U.S.-China trade war and Brexit?

Segars: The immediate impact has been relatively low. The semiconductor industry has taken a dip, because so many companies supply into China. Generally, when downturns happen or inventory issues occur, we tend not to worry about it too much. A lot of our customers are working on things that are several years out. Our revenues today relate to things we designed a long time ago, so our costs and our revenues are completely disconnected. If you believe AI, 5G, IoT and autonomous vehicles are going to be the drivers of growth in the future, then through good times and bad, we need to be investing in the roadmap for what we believe our partners are going to need.

SE: China is now developing a separate supply chain. Does that cause problems or create new opportunities?

Segars: It’s too early to tell. Global supply chains can work really well, and work moves to the most efficient place where it can get done. In a more protectionist world, that naturally leads you to inefficiencies because people will replicate structures in other parts of the world if they feel they can’t rely on something. It’s self-preservation.

SE: It’s also a hidden cost, right?

Segars: Yes. If production gets more expensive, the consumer pays for it.

SE: Does this slow down investment, as well?

Segars: Maybe, but it also can lead to a surge in investment. If a market is tied to a certain thing, then they want to invest in that thing. Instead of one winner, there can be multiple regional winners.

SE: But it does make it harder to balance inventory, right?

Segars: Yes, it will make supply-chain management more complex, and ultimately more expensive, if the world ultimately does dislocate into different regions. I hope it does not, but we have built our engineering resources in a disaggregated way. The U.K. is the largest single site. We’ve looked to build engineering centers in key talent pools around the world, including the U.S., Israel, Taiwan and multiple countries in Europe. We’re not all concentrated in one place.

SE: We’ve talked about security in the past, but we have more data online and more breaches. What’s the solution?

Segars: It’s a challenge, and it’s only going to get worse. We are still in the very early stages of deployment of IoT. There is more data and there are more ways of getting into a network, and both of those will explode in coming years. We’ve put a lot of our focus in trying to get that right from the start. We have things like PSA (Platform Security Architecture) even for small microcontrollers, which people view as a trivial simple device. It will have bad security if you look at it as a trivial simple device. PSA has been trying to address that issue, but we aren’t going to solve this on our own. You need industry collaboration, which is why we’ve been pushing PSA. We helped form a committee in the GSA to look at security. And at the other end of the spectrum, there’s work on long-range architectures for the cloud to enhance security there. There’s a consortium coming together to try that technology. There also was a new Linux Foundation consortium that launched a few years ago called Confidential Computing, which is about privacy and unique security. The good news is there are more and more big names looking at security. The bad news is the perception that security is someone else’s problem and customers are not willing to change for it. That is slowly starting to change.

SE: There are so many devices that are not secure in addition to those that are secure that it raises questions about whether anything or anyone is vulnerable.

Segars: There is a program in PSA that does start to do ratings for devices. Hopefully that will be rolled into something the industry can take hold of. The fact that there are insecure devices out there won’t go away for a very long time. There also will be stick-driven manual cars out there for a very long time. All of this has to co-exist. To manage insecure devices, one role of the edge device will be to interrogate anything that is connected to it. So the role of the gateway in assessing what’s connected to it and monitoring what’s going on becomes a much bigger component of the overall security.

SE: Security and privacy also vary by country. So China will not let any data reside in a car. Europe is concerned about privacy. And the U.S. is pretty much laissez-faire. How do you manage that?

Segars: At the moment, it’s the Wild West in the U.S. when it comes to privacy. But that’s changing. There are discussions about some form of privacy regulation. In Washington, people are saying there will be a total disaster if there are 50 versions of privacy because nobody will be able to get anything done.

SE: How about the supply chain security?

Segars: Chips may come from different fabs that may or may not be owned by the same company. If the chiplet model takes off, there will have to be some standardization about how dies get shuffled around the industry and different standards. But the semiconductor industry is good at collaborating around standards that make sense.

SE: There are a lot of offshoots of markets coming. What looks interesting to you?

Segars: We’re trying to look broadly about what you need to build AI systems that can understand natural language. What’s good is that the cost of technology keeps coming down. We don’t have to pick specific end markets. It’s trying to understand the computing building blocks.