Executive Insight: Jack Harding

SE: What’s worrying you these days?

Harding: One thing that bothers me is the cost of chip development on a per-chip basis. We seduce ourselves into thinking everything is wonderful because the cost per transistor is dropping in chunks. Gate costs are going down at every node. If you look at the secular trend, we’ve done a pretty good job putting a lot of stuff in a small space. In my business, I’ve seen the well-defined and explicit definition of NRE, and NRE for a 28nm chip will run $5 million to $10 million. As we look at 20nm, it’s $8 million to $12 million, and with finFET they’re $15 million. This is tooling, IP, design work. I’m worried that what’s happening is we’re reducing the market down to the 8 or 10 guys who can afford it and the 4 or 5 guys who can make it.

SE: There’s a lot of pressure to stay at existing nodes and roll backward, though, right?

Harding: We’re seeing the beginning of that. We’re being told, at least anecdotally, that people are going to stay at 28nm much longer than at previous nodes. We are heartened by the level of technology needed for the Internet of Things—those trillion parts that have been on the horizon for 10 years. These are relatively simple parts. They’re not consumer-like, but sub-consumer-like. The 10 devices on my desk may be enabled by sensors and actuators that become relevant. That’s the other eternal promise of 2.5D.

SE: What’s your current view of that approach?

Harding: I see great promise there. We’ve got hardware back. We’re demonstrating some very interesting technology already. I thought we would have more problems from a supply chain perspective, but those things are working themselves out.

SE: So are you optimistic or pessimistic about the future?

Harding: When you look around you’ve got the cost of the classic SoC going through the roof along with the complexity and you have to ask who’s going to ask for it and who’s going to make it? But there are also trillions of sensor/actuator parts, improved architectures around existing nodes, and improved packaging techniques that give us die re-usability. I measure goodness and badness versus what they looked like 10 years ago versus today. This industry has never dropped the ball, and I believe we will continue to do great things. But there will be different curves than what we have today, and we will have to adjust and cope with that.

SE: Who do you see actually moving to the next nodes in terms of numbers and types of companies?

Harding: It’s the biggest and baddest guys. It’s the guys with the P&L to support making these kinds of bets. They can write a check for $10 million or $15 million to try it. They’re going to be the top 10 OEMs. There will be some fabless semiconductor guys, too, who believe they have a sufficient breakthrough with an aggregation of multiple capabilities into one chip, maybe for a cell phone or tablet. There are 8 or 10 of those. And after that, everyone runs for the hills.

SE: Do you see that potentially as a platform for a 2.5D and 3D chip, where maybe that’s the logic piece that runs everything?

Harding: Those 8 or 10 companies will do one or two chips a year. But for everyone else, they may have a huge memory stack and SerDes they can put into a 2.5D stack. They may add logic with 16nm. Or you may have guys who say they’re not looking to do high-performance computing and they’re happy with FD-SOI with comparable performance at less power and not having to pay the big bills for finFETs. The value-add will come from those companies or individuals who have the architectural mindset to integrate at the die level, not at the transistor or block level. You have to understand the power from die down to transistor level, that will be the required skill set. You think about it as a die, but you need to understand what’s going on inside this chip at the transistor level to make sure it works. It’s going to change the orientation from the horizontal view of laying out transistors to a vertical view of different form factors that have to play together in a total solution.

SE: Does this open up the market to new players, or is it still the same players?

Harding: That’s the $64,000 question. For all the other buckets we have answers, but for this one we don’t know. It seems that if you’ve got a module or PCB capability and you’re a company that thinks in those terms, the MCM/2.5D approach is a natural one. Those people should become the value leaders because they understand the physical implementation. I would like to think that will lead to a reusability that is very attractive—the mix and match—where you’ll see exchanges growing and partnerships. That will drive more parts and more players. But you’re talking about cooperation between companies rather than within companies, and that takes longer. I’m optimistic this will happen, but it’s going to require strange bedfellows and relationships and contracts. This is not unlike when we started the fabless business model 14 years ago, where we said we would rationalize all these players. We will rationalize all the partnerships that have to happen to optimize a 2.5D solution.

SE: So isn’t part of your role relationship management?

Harding: Yes. Instead of supplying wafers you supply good die. There is a relationship management function, but it’s not all that different from what the ASIC fabless model is today. You have 50 different suppliers that we mix and match for hundreds of customers in logical and predictable ways. That can happen at the die level for 2.5D. But I’m also hesitant on timing because the leadership inside of companies will have to shift from the ASIC guy to the one who thinks in terms of packaging and form factor and transistor management.

SE: Won’t some of this be driven by the ability to develop at 90nm or 130nm where it’s less expensive?

Harding: That’s precisely the impetus. People will go into partnerships and relationships with other companies for just that reason. It’s like today where you have a certain class of chip, and if you can take that die and stack memory over here and here’s another piece from over here and another one from here—that’s what has been happening informally for a while. That will evolve into a marketplace where people can figure out what’s available—almost like a consignment shop for old technology. But if we know for sure that will increase the number of chips in the market at the 2.5D or MCM level, then we also know for sure that the number of SoC starts will drop.

SE: Does that kill off the superchip concept?

Harding: There’s an interesting bifurcation going on. Let’s say I bite the bullet in a 16nm finFET chip and I pay $12 million or more for NRE. I have the verification behind that, so it grows to $150 million. But now I’ve got this very small die that is relatively inexpensive. The unit cost for that die is quite low, considering what it’s doing. That chip went from being an unattainable chip by virtue of the NRE to a very easily re-used chip for a 2.5D application. It may cost a couple bucks. But if you pay through the nose up front, you might be able to sell it six months later for a cost that is cheap enough to easily go into a 2.5D package. There might be a secondary market for the SoC.

SE: So it’s a very long tail with high volume and low cost?

Harding: Yes. And what if from now on you say, ‘To justify the NRE you must sell the die into the 2.5D market?’ That could be an expression we use three years from now. So how do you do that? Maybe you turn off different chunks of logic to reduce power, to protect IP and to improve functionality. You might make architectural decisions that, after you’re done selling this chip to the No. 1 tablet company in the world, then you will sell this chip—which is almost configurable—to the 2.5D after-market to give them access to technology they wouldn’t otherwise be able to get. You can protect some of your secret sauce. Those are the discussions we may have to enable this secondary market.

SE: That’s a radical shift in the SoC business model.

Harding: Yes. The model changes. The architecture changes. But we also may look back in three or four years and find out this idea was totally wrong. It wouldn’t be the first time. On the other hand, you might find companies paying even more to get a chip out the door because they know they have a secondary market that justifies the initial investment. There will be different thoughts on this. This is an ROI discussion about how it impacts the chip architecture. That’s a horizontal discussion. The vertical discussion will involve the board-level guys who will be talking about form factor and function problems they have on the printed circuit board. No one is quite sure how that will change.

SE: How real do you think the Internet of Things is?

Harding: I do think it’s real. I expect to see tens of billions of things getting shipped. I remember 15 years ago when futurist Paul Saffo gave a presentation. He used a different term, but he described it perfectly. I have a friend who acts as a beta site for a front-yard sprinkler system. He monitors his lawn from anywhere in the world. Every sprinkler head has a sensor in it. It’s happening, and it’s not going to stop. When you see 5 billion of the 6 billion people in the world getting into technology, they’re going to leapfrog anything with a wire-based infrastructure and go straight to sensor-based actuator technologies. I believe large companies will do top-down deployment of hundreds of SKUs (stock-keeping units) of IoT-class chips that will make them quite successful, but to me that’s not the most exciting market. The IoT market, given the relatively low cost and complexity of these chips, will be served equally by open-source semiconductor design. Just like you see people doing open-source software, where someone is writing code in India based upon code written in San Jose, you’re going to see the same thing as it relates to Internet of Things.

SE: How important is security in all of this?

Harding: It depends upon the area. If someone turns on my sprinkler when I’m gone, that’s not an issue. If someone unlocks my door, that’s a problem. The marketplace that the broader semiconductor community needs to support and facilitate is one where you share circuit-level ideas that allow people to design things faster with quicker turnaround and in a low-cost way.

SE: But there are two sides of the IoT, right? One is simple, the other is rather complex.

Harding: Yes, the complex one is the infrastructure piece. We’re closely monitoring the infrastructure elements of that system and have set up a system to automate much of that through online quotes. We’ve done the back end from test chip to tapeout to production—we’ve automated the whole thing—and now we’re working on front-end architectural capabilities to let people make tradeoffs before they commit to the heavy investment, whether it’s $10 million or $2 million.

SE: What do you think about adding in software?

Harding: It’s not part of our business right now. The software business is 10 times or maybe even 100 times the size of semis right now. It’s becoming more and more the domain of our customers. Our intent is to exploit the movement away from hardware to greater investment in software—letting our customers do that while we take care of the silicon. The secret sauce seems to be software, and that’s what they’re chasing. We’re not application guys here. We deliver hardware in the most cost-effective way—with a robust, transparent, easy to use infrastructure. The big question is whether we will enable more people to make chips if we make it easy enough.

SE: Have you seen any shift in who’s using your automated tools and placing orders versus in the past when you didn’t have them?

Harding: Our online engagements are coming from more than 40 different countries. Prior to that, we were selling into fewer than 15 countries. So we’ve tripled the geographic coverage and we’re dealing with people we’ve never talked to and in countries where we never set foot.