Experts at the table, part 2: Why some companies are moving to finFETs while others are staying put; the value and challenges of subystems; developing IP for a splintering market.
Semiconductor Engineering sat down with Chris Rowen, CTO of Cadence‘s IP group; Rob Aitken, an ARM fellow; Patrick Soheili, vice president of product management and corporate development at eSilicon; Navraj Nandra, senior director of marketing for DesignWare analog and mixed-signal IP at Synopsys; and Kurt Shuler, vice president of marketing at Arteris. What follows are excerpts of that conversation. (To view part 1, click here.)
SE: How big of a concern is the consolidation underway in a variety of markets, including automotive electronics?
Shuler: I’d be worried if it was just a cost-based thing, but that’s not what’s happened. They’re buying companies than their core business but they want to expand beyond that core business. For us it makes it easier. You have two sets of customers. The more users you have in a company, the better. The cost of sales is easier within an existing customer.
SE: It seems as if the IP industry is shifting in a couple of directions. We have the big IP vendors selling more custom IP. Then we have the guys going for a socket who may be going for a narrower set of characterization parameters. But no one is characterizing it for a broad set of uses, right?
Aitken: That’s a good way to look at it. There’s a segment of the market that wants to differentiate on whatever their secret sauce is and they really want to get to the market quickly, so they’re looking for systems or subsystems that are pre-constructed, pre-characterized or pre-defined. They can take them and use them quickly. Even if what’s being provided is RTL, that RTL has been validated in silicon. These customers are looking to add onto what’s already there. That’s why the characterization problem can change over time. Instead of characterizing a million separate blocks you can characterize one system and say, ‘Here’s how it works.’
Shuler: In our market research we found there is a generally accepted architecture. Then, once you have that, you look for who are the players that can provide the individual pieces. The only place that doesn’t work is in the IoT.
Rowen: Clear-minded architects gravitate toward saying they’re going to have clear differentiation in A, B, C and D, and they’re going to make their lives as easy as possible in E, F and G with whatever is the standardized way of doing it. They want to minimize risk there so they can maximize risk on other matters. The better the team is at understanding their engineering tasks, the better it works. That is different from the past, when the old-line semiconductor companies saw their job as innovating at everything. They developed their own IP internally.
Nandra: It was IP and tools.
Rowen: Yes. But when you take that approach are you really differentiating or just guaranteeing full employment?
Nandra: Depending on the particular market segment, that characterization is very specific. We’re seeing unique IP needs for automotive, for example. So it may be the same IP block, but it needs to be characterized for that particular market. That’s another level of sophistication I’m seeing with the IP companies. You’re no longer selling a generic block that fits into every market segment. You’re designing it for an extended temperature range, for example, with some special PVTs for environmental conditions. That would be overkill for a consumer market.
Rowen: That’s a very important trend and it’s closely tied to the subsystem question. Subsystems, by definition, are much more narrowly focused than the components that go into those subsystems. The selection of components makes it much more targeted. It’s a particular combination of things that makes it unique to a specific application. And therein lie both the value and the challenges of doing subsystems. They are for narrower markets but they’re more valuable combinations.
Nandra: For the IP vendor, the question is how far do you go with your subsystem? Do you just prove the concept because it’s not something the customer necessarily will need for their application? There are a couple approaches here.
Soheili: Seventy years ago when Procter & Gamble came out with Tide, there was one kind of Tide. Now there are many types. This super-segmentation of the market happens in part because they want to grow their market, and in order to grab multiple people they have to differentiate the product. The differentiation got pushed to the manufacturer and that’s where it stayed for a long time. The activity in our market is being pushed to the IP vendors because we want to sell our product and differentiate. So a lot of this is self-inflicted.
Nandra: You see that with the foundries. Every few months we’re seeing a new version of finFET technology or an improved 28nm. That’s an attempt to grab more of the market. But if you talk to our customers, they’re really confused. Do they do HPC, HPC-plus or some other version? It’s a laundry list.
Aitken: One reason why they make so many types of Tide is that every shelf occupied by Tide isn’t occupied by Tide’s competitors.
Nandra: And you’re causing a conversation about ‘my foundry with my technology.’
Rowen: And we assume there are challenges for each version.
Nandra: With IP, we have to do the characterization for each variant of the process.
SE: One of the trends we’re seeing is that the market is splintering between those heading to 7nm and beyond, those sitting comfortably at 28nm, and those staying much farther behind. What does that do for IP development?
Soheili: It’s the same as it always was, but the forces motivating who’s going in what direction are harder and harsher. If you have to integrate, and you’re a GPU guy, you have to go to 7nm when it’s available, no matter how much the mask set or how much the IP costs. If the ecosystem isn’t there, you do it yourself. The guy who can get away with 180nm is going to stay at 180. If they’re doing an RF device, there’s no reason to move to finFETs. But there’s also a small number of people who are sitting the middle, too. They could push this way and get these benefits, or stay here and not get those benefits. You take your risks here, but not there.
Aitken: There’s been a tendency over time for a bunch of people to sit at 180nm or 150nm and be happy there. And there’s another group rolling forward with Moore’s Law. What we’ve seen recently is that wave two in finFETs is really close to wave one. People you wouldn’t have expected to go to finFETs are going there. That’s an interesting trend.
SE: Is that confused by the fact that chipmakers didn’t gain much by moving to 20nm?
Aitken: 20nm stalled out, but 28nm looked like and still looks like it’s going to be a very long node. That’s why there are 120 flavors of 28nm. Everybody’s trying to get the right one in there. But a lot of companies we thought would stay at 28nm are now going to 16nm and 14nm because they’re not as scary as they looked at first glance and because their competition is moving there. Once their competition goes, everyone has to go.
Rowen: The last thing a vendor wants to do is be fighting with a 28nm sword in a 14nm finFET battle. It makes it really tough.
Shuler: What we are seeing, though, when you look at mobile phones and application processors for digital basebands is that there are one or two products in the product line that are going to 16/14nm, but the herd has been culled because of cost. It doesn’t make sense to move the whole product line down there. 28nm from a cost standpoint is going to be mainstream for the next four or five years.
Aitken: If you look at foundry sales, 28nm ramped really fast and is likely to be there for awhile.
Soheili: We’re quoting a bunch of ASICs where we could go to finFETs, but they’re staying at 28nm. The scaling either convinces you to move or it doesn’t. If it doesn’t, why pay that much more and take that much more risk just to be with the herd?
Rowen: The market is splitting into two or three areas. First, there’s 55nm, which is the new 180. That’s how it’s positioned. There’s 28nm, which is for people who want modern technology without the risk, and then there are the bleeding edge guys. There are a surprising number of those, and they’re people with deep pockets so it’s an important market. But how does it affect the IP business? There are some very important second-order process effects on soft IP. And there are a huge number of first-order effects for hard IP. This choices are where do you invest, which customers do you listen to most, and of all the opportunities which ones do you pick—and there is a huge demand for optimized hard IP on every new node. There’s a joint economic decision about what do people choose.
Aitken: Some of the things you think are second order are actually first order for soft IP, like can you do voltage and frequency scaling? At 16 and 14nm, yes you can. At 28nm, you can scale from here to there. The overall scope of a product to fill a performance band is much greater with a finFET than without one.
Rowen: If it’s developed for 16nm, you find you can easily roll back a node. We’ve done a lot of work on near-threshold finFETs on small processing engines, and while the libraries aren’t characterized as we want, we could take the whole processor and run it through SPICE. It works great.