The Good And Bad Of Models

By Ann Steffora Mutschler
Driven by fierce competition and the fact that socket decisions are made long before silicon is manufactured, semiconductor companies today ship models and virtual prototypes to their OEMs very early in hopes of locking in the socket. Admittedly, this has been happening for some time, but due to complexity and the need for flexibility of models and virtual platforms, the semiconductor suppliers want to easily distribute their models for anyone to use.

“The tricky part is who is willing to discuss it because a lot of people have been doing it. We’ve had a number of press releases and public statements, and we have a deal in place with Freescale to distribute some of their IP to their lead customers and new customers who are trying to use their products before they’re released,” noted Jon McDonald, technical marketing engineer for the design and creation business at Mentor Graphics. “For Freescale and ARM—and ARM was one of the biggest drivers of this trend—what they are realizing is that they can’t win a socket and they’re not going to sell their silicon unless they have the model. If they have the model and they can get designed in up front before the silicon is even available—before there’s even competition for the socket —they can get locked in. That’s probably not so much ARM’s motivation, but for Freescale, I think that’s a big part of it. They want to have the models that they can essentially lock up the socket.”

Johannes Stahl, director of product marketing for system-level solutions at Synopsys agreed. “We have reached an inflection point where it becomes important for the semiconductor vendors to ship these virtual prototypes to their OEMs.”

This is now beyond the experimental stage with companies such as Altera, Renesas, Infineon and many others looking to provide virtual prototypes to their customers. “It’s going from the pure internal uses such as what ST has done to speed up internal development now into the market deployment and helping the end customers—the OEMs—to be successful with the silicon,” Stahl observed.

Frank Schirrmeister, group director for product marketing of the System Development Suite at Cadence sees two trends at work. “One is certainly the interaction between the semiconductor and the OEM and sending something to the OEM. There’s also the adverse trend, which is the semiconductor providers have to provide more and more software themselves. There’s really the dynamic of, ‘Do I send it outside or am I forced anyway as a semiconductor provider to hire more software developers to provide more software myself, in which case I’ll ship this virtual prototype to the next user and to the next building in my company, or the next office.’”

This raises the question of how to provide models at the same time the OEMs are driving their requirements downwards. As such, the amount of responsibility is moving up in the supply chain.

While this is not new it has accelerated, he said. “Think about the following stack from IP going into the chip, then on the chip there are various levels of hardware abstraction layers, operating system drivers, middleware, etc. Then this all goes into the system, and the board is involved somewhere. All of that together is the system, which the applications are written on. The move always has been a struggle but the OEM is definitely controlling more of it.”

In addition, it’s actually moving beyond the OEM. “Do you think in the wireless space the OEM does anything? Not really,” Schirrmeister asserted. It’s the network provider that says what they want.

Case in point, at one time Verizon offered NFL footage through its network on its phones. That translated to changes throughout the supply chain from the networks needing to be upgraded to have the right bandwidth, on down. “That also happened when Apple came out with the iPhone; the bandwidth went to its knees in the beginning. The phones then have to be upgraded. It all ripples down,” he said.

On the other side, this translates to the semiconductor IP providers and everybody else having more responsibility. “The IP guys are moving toward subsystems, the semis have to do more software and provide software themselves, and have all the ports ready,” Schirrmeister said. “It’s a back-and-forth like a wrestling match and it goes in both directions. You see the OEMs defining requirements down into the chain and the semis are providing executable items like virtual platforms, virtual prototypes upward. And that goes into both directions. The OEM may take a virtual platform and performance annotator and say, ‘Here’s the performance I need from your chip and if you don’t get that performance, I can’t use your chip.’ Then on the other hand, there’s the virtual platform that demonstrates from the semiconductor guy upward what’s going on that actually gives me the software development, for one, but then the power and performance analysis aspect that can be annotated and become part of it.”

As it stands today, more and more OEMs are not making decisions on physical silicon. It’s just too late, said Mentor’s McDonald. “There’s no way to get it to market in time if you don’t make the decision on what the silicon’s going to be until the silicon’s available. You’d be starting out so far behind the goal line that you just can’t get there.”

As far as how this evolves over time, the problem the semis have is they want their models to be available, he said. “They want the models to be widely available, freely used and instantiated,” McDonald said. “They don’t want to have a huge burden in distributing and supporting those models. They don’t have the infrastructure in place or software or EDA-style distribution. What they want to see happen is that they can create models, make them available and then anyone can use them.”

Some of this also benefits from the evolution of standards such as SystemC and TLM 2.0.

“We can take models from a Tensilica with very little interaction from our customers’ point of view and instantiate that model with other models from other sources today,” said McDonald. “So having the standards and being able to create models that are not tied to a single environment is ultimately going to be very important because if your models are all tied to a single environment, again, you’ve got compatibility issues. You’ve got to support different environments. What if I want to use models from different environments? When things are not compatible, that causes problems for the customers and I don’t think that’s what the silicon providers want. They want to put out a model that everyone can use.”