Experts at the Table, part 2: Why there is less finger pointing today, and why subsystems are still not pervasive and probably never will be.
Semiconductor Engineering sat down to talk about the future of IP with Rob Aitken, R&D fellow at ARM; Mike Gianfagna, vice president of marketing at eSilicon; Judd Heape, vice president of product applications at Apical; and Bernard Murphy, an independent industry consultant. What follows are excerpts of that discussion, which was held in front of a live audience at the recent Silicon Valley IP Users Conference. To view part one, click here.
SE: What kinds of problems are showing up as the industry relies increasingly on third-party IP?
Gianfagna: One of the real issues here is choice—what libraries to synthesize against, what number tracks in the library, what Vt mix, what process option, what foundry. The list goes on and on. And those dials have fundamental impact on how hard it’s going to be to debug the design in the first place. You want to try to ring out as much uncertainty as possible so the number of bugs you’re chasing is small. There’s some analysis that can be done up front to help this.
Aitken: Those kinds of bugs are the timing closure bugs as opposed to the functional bugs.
Gianfagna: Exactly. Timing closure, power closure—you name it.
Aitken: Signoff bugs.
Gianfagna: People come to us all the time and say they want to design a chip, but they’re 20% over power. What do you do about it?
Aitken: We were in a similar situation when we came up with the whole POP IP concept to deal with that. http://www.arm.com/products/physical-ip/processor-optimization-pack/ We say the core goes X gigaghertz, but when you implement it you find it goes much slower than that. So we provide the recipe and if it goes faster, that’s great for you.
Gianfagna: Yes, that early analysis goes a long way to smoothing the way later on. It’s never smooth, but it can reduce the number of functional bugs you’re chasing, get rid of the timing/power/performance closure issues—or at least make it easier to get there.
SE: How much finger pointing is going on these days compared to five years ago? Chips are getting more complex, but at the same time the IP is getting better.
Heape: We really try to work with our customers to make sure things are done correctly. When things don’t work it’s typically because of one of two things. First, the spec is misinterpreted or ignored. Or second, they were improperly documented from the IP provider side. Those are the things we run into most often.
Murphy: There is a certain amount of self-inflicted thrashing that goes on. At the end of the day the chips come out and they work. But in the middle it gets kind of messy. And some of that mess is self-inflicted. As an industry, we still don’t agree on quality standards. We’ve taken multiple passes at this. Maybe it’s the kind of thing that doesn’t lend itself to a standard. Maybe it will need to be a de facto standard or we have to publish, ‘This is how I did my checks.’ But we are definitely getting thrashing as a result of not having alignment on these things.
SE: But that standard would need to change, right? We now have things that weren’t considered important in the past. Power is one of those. And security is coming up fast.
Murphy: You evolve the standard to deal with that. There are baselines that don’t change. There are constraints checks. Maybe those are more debatable now that timing is becoming more complicated. But there are regions where we could standardize. CDC is a perfect example, where we are sacrificing standards in favor of product purchases. If I use product X from one company, product Y from another, and product Z from another, and I’ve checked my IP against these, I’m not going to check it against anything else. ‘I didn’t find any clock domain crossing issues, but if your product did, that’s your problem.’ That’s crazy.
Aitken: But what’s the alternative? We’ve moved to a de facto Yelp type of mechanism where there is a community someplace that says, ‘This stuff seems to work so you can use it, but this stuff is kind of sketchy so you may not want it.’ That works well if you’re building something. It adds a barrier for entry if you’re trying to get in.
Gianfagna: There’s definitely less finger pointing, but it’s not because the IP standards are better. As the chips get more complex, particularly at 14nm and below, the number of companies that can do these chips is shrinking. The people left standing are smarter and smarter consumers. They know what questions to ask, and they get better at picking the IP that works together because they have more mileage on it. That reduces the finger pointing. It doesn’t reduce the overall number of problems, but it does reduce that problem.
SE: Does it reduce the number of companies that can play in this world, as well?
Aitken: No, but the landscape does change. There is a small collection of people working at 16/14nm, and they really do know what they’re doing. Suppose you want to get into that business. We cannot reinvent all the background knowledge that exists in all those other people. It’s impossible. But what we could do is attempt to procure as much of that as possible in the form of a subsystem or some other pre-designed thing. These are starting to come into existence. That’s one way to approach it. ‘I’m not going to try to tweak this. I’m going to take something that somebody already made work, and whatever secret sauce we used to start our company we’ll use to differentiate, not the vast bulk of our chip.’
SE: If you believed what was being said five years ago, everything today would be a subsystem. The reality is it’s gone the other way. What’s changed to make that happen?
Aitken: Subsystems are a challenge to design. If you say you’re going to build a subsystem, you create something in PowerPoint and it looks spectacular. Then you realize there’s no PowerPoint-to-GDS converter. But you rapidly get into a whole bunch of details like where do you find one piece of IP that no one sells anymore. It’s the nitty gritty of finishing it that’s hard. And then, once you finish it, who wants it and for what? Nobody wants a hardened subsystem. If you provide a semi-modifiable one, more people want it but that one is hard to verify. Trying to find a middle ground and what market it exists for is an interesting challenge.
Murphy: I agree. It’s more driven by economics in the market than design. What is it you want to build as a subsystem that would meet a wide range of needs? That’s a lot less clear than it used to be.
Gianfagna: Everyone uses the term subsystem to refer to the new macro building block, where you only need two or three of them and you’re done. That’s not happening. Are IP blocks getting more complex? Yes. The IP block of today is the full chip of yesterday. But the market homogenization isn’t happening. Everybody doesn’t want the same technology because then there’s no differentiation. At some level there’s customization, and that’s how this one wins against that one. That never goes away. The blocks do get better, but subsystems imply there is a sea change in the way chips get built, and we don’t see that.
Murphy: At one point we all thought applications processors were going to drive the way chips were built. That was a beautiful picture for awhile until the model began falling apart.
Heape: We deliver imaging subsystems, but in reality no subsystem we’ve delivered in the past six years has been the same. There is always a little customization. It may be as simple as an interface change. Or it may involve a different imaging sensor. We’ve tried to do three standard products, but we’ve never sold them. We’re always changing them.
Aitken: If you decided to become a subsystem provider, it would make more sense internally to have some process for putting things together and validating them, and then releasing them, rather than either having a pre-defined object you try to sell or a tool.
Heape: That’s what we do. Our builds are scripted. So we can build a pipeline and move in and out blocks.
Gianfagna: Automated customization is a good way to make money.
Murphy: Systems that may have broad appeal may be more like arrays of sensors than digital logic.
Aitken: That’s the claim we’ve made in ‘IoT Land.’ There are enough people in the long tail of the universe who don’t know how to build something or can’t be bothered to understand how one of these subsystems fits together. They would be interested if stuff was put together.
Murphy: Yes, the amplifiers and sensors and analog pieces that people don’t know how to do.