Experts At The Table: Which Comes First?
By Ed Sperling
System-Level Design sat down to discuss hardware and software priorities with Neil Hand, group director for marketing for Cadence’s SoC realization; Johannes Stahl, director of product marketing for system-level solutions at Synopsys; Prasad Subramaniam, vice president of design technology at eSilicon; and Bernard Murphy, CTO at Atrenta. What follows are excerpts of that conversation.
SLD: Which comes first in a design? Is it the hardware, the software, or both?
Stahl: What comes first is the user experience. So you can say it’s the software, but no user experience can happen without an efficient hardware platform. But the hardware also needs to have all the components to support that user experience.
Hand: That’s the big change that’s happened recently. It used to be hardware first, and then software built on top of it. Now you have to ask, ‘What is it we’re trying to build?’ While the software drives the hardware requirements, the hardware has to enable differentiation. Otherwise everyone is competing on a commodity basis and that’s not a good thing.
Murphy: The reality is that semiconductor companies aren’t creating platforms from scratch. There’s a lot of IP value in things they’ve built up over the years—DSPs and Codecs, for example. That has to fit in with the software. If you have a multiprocessor platform, one of the eternal challenges is how you verify that you won’t have contention problems or resource starvation given the hardware you’ve chosen. You need to ensure you’re not going to get lockups, so I think it comes from both directions, not just software.
Subramaniam: I do agree the end application drives the design. I wouldn’t go so far as to say it defines the platform, though. Once you define the platform you’re defining a common denominator. If you want to differentiate yourself obviously you need something unique for your application. The end application drives the hardware design. The hardware may have some unique components even if it’s just the processor core. You also have to consider what’s going to be built in hardware and what’s going to be built in software. Then those parts that are built in hardware will be custom-tailored for the application. So it’s a combination of both, with the application dominantly driving the design.
Stahl: But the question is what you mean by ‘the application.’ If I’m designing a smart phone or a tablet there are hundreds of applications. There is no single application that will drive a design. We are talking about a myriad of applications, and we have to make sure we develop an architecture that can support all of these.
Subramaniam: While that’s true, if you look at the underlying implementation of the smart phone or the tablet it’s a unique experience—how you do the graphics and how you display Web pages. These all have to be done independent of the user application. You look at the underlying features and you focus on what you need to do to implement those features most efficiently.
SLD: So do we have to move this discussion up a level of abstraction?
Hand: If you look at DTV (digital televisions) these days, even they have applications running on them. That doesn’t change the underlying requirements of the hardware, but there will be a commonality in whether you’re going to be running 2D games or 3D games. The requirements for a tablet are very different than a digital TV even though both run YouTube. When we say apps we’re really talking about a class of applications.
Murphy: But there is a hardware component to all of this that the apps can’t create on their own. There are emerging capabilities in near-field communications that enable a whole new realm of user experience. That’s why it is a chicken-and-egg problem.
Stahl: At the end of the day it comes down to what are the critical things for the platform. The critical things are the high-performance hardware that really can affect your power consumption and drive your architectural decisions. What we see in the market is that customers spend a lot of time going through all the high-performance scenarios and evaluate them thoroughly before making a final choice. How many processors and accelerators will you need? This is done very early in the process, and typically in collaboration with the OEMs, the user and the semiconductor vendor.
Hand: What we’re seeing is, rather than higher performance, what is the required performance? They want to get to the lowest technology required to support that. What is the slowest memory you can get away with for streaming video? What is the cheapest implementation of the on-chip architecture? It’s not just about getting the most out of a chip. It’s also about lowering the cost for the bill of materials. For some companies the differentiation will be price. For Apple, the differentiation is the user experience. And for Google’s Android, it’s about how many cool new features can be supported with each one. Differentiation drives what will happen on the hardware and the software level.
Murphy: Another consideration on top of PPA (power-performance-area) is what’s being called PPAX, where X is the schedule. You don’t necessarily optimize for power, performance and area if it’s going to take too long to do it. There’s a compromise in there, and you see that on the software side with Android. People aren’t using all the capabilities of Android because it’s not worth it in terms of time.
Hand: Especially when the product lifecycle for an Android platform is almost measured in weeks now.
SLD: So what is good enough these days? Is that changing?
Stahl: What is not good enough is when you have a smart phone where you have to change the battery every hour.
SLD: But we do put up with charging it every day, right?
Subramaniam: Good enough is a moving target. What is good enough today is not good enough tomorrow if someone else comes up with a better solution. That’s why it’s dangerous to design to good enough. You have to commit yourself and say, ‘This is the line I’m drawing.’ Then you have to design it in the most efficient way to meet this objective. But someone else may have drawn the line further ahead.
Hand: There’s a tradeoff, because you’ve got to get it to that line on schedule. Otherwise you keep making it better and better and you never get done. Everyone draws a line. It’s a short lifespan for smart phones. It’s a different story for automotive. People will put more into the design because it’s a longer lifespan.
SLD: What’s the most important factor in designing an SoC for the consumer market? Is it technology, cost or time to market?
Subramaniam: Technology is probably last. The most important thing is time to market and then the price. You do have the luxury of turning them over very quickly. But you’ve also got to hit market windows, like the holiday season. Schedule is paramount. You have to design to meet that schedule. And then, because it’s a highly competitive market, you have to deliver it for an acceptable price point. Delivering on time and for the lowest price is the technology of choice.
Hand: Technology depends on where you’re competing. At the low end, schedule and price will rule. If you’re trying to compete with the latest ‘iDevice,’ technology matters more. That’s part of the challenge of the SoC makers. It’s very difficult if they’re going to target a wide range of applications. If they’re covering DTV, automotive and consumer with a single device, that becomes difficult because there are a lot of tradeoffs there.
Murphy: Automotive in particular is challenging.
Hand: Increasingly in automotive it’s about the entertainment system.
Stahl: There also is government regulation involved in automotive. The entire industry is scared of a 15% increase in automotive design prices because of safety requirements.
Murphy: They’re equally scared of the Toyota effect.
SLD: Is it necessary, then, to push to the next node of Moore’s Law?
Subramaniam: It’s a tough decision. If you look at the transition from 65nm to 40nm, and then from 40nm to 28nm, the transition has slowed down. Previous changes in technology happened much faster. That’s because you have to look at the investment. People will stay at existing technology nodes longer.
Hand: It goes to the question of what’s good enough. If you can get the job done at a trailing node within the power and performance requirements, then why go to the leading-edge node unless you have the volumes to justify it.
Subramaniam: The product leader in that market segment will probably want to go to the next node, and they have the confidence level that they’ll be able to recover their investment. The other guys are going to have to figure out whether they can match that, and more often than not it will take them longer.
