Betting On Subsystems

Bundling and pre-verifying IP blocks makes sense on paper, but how far will this trend continue?


By Ed Sperling
One of the consistent trends among successful companies, particularly in well-established industries, is that over time labor becomes specialized. No one can do everything well, and the more complex the systems the more pieces have to be outsourced.

This creates immediate benefits for companies putting together the overall systems. They can focus on designs and doing what they do best. And it creates new problems, because standards have to be created and updated so the pieces fit together exactly as planned. This is particularly difficult in complex SoCs, so rather than breaking pieces down to their most basic component level, they are partially re-aggregated into subsystems.

The automotive industry went through a similar transition. In the early 1900s, when cars were still basic and the industry was new, all the components came from the same companies. By the 1950s, there were suppliers of most of the components inside of cars, but gradually those aggregated into radios, seats, cooling systems, and more recently there are suppliers for wiring harnesses and the electronics that are replacing mechanical components.

From IP to subsystems
Subsystems are the likely follow-on to IP blocks inside of SoCs for several reasons. First, it gives vendors a way to capitalize on their success in the IP market—and to distance themselves from those companies that failed. Early adopters of third-party IP learned quickly that just because one IP block is priced lower than another doesn’t mean it will end up costing less in the long run.

It costs a lot, in terms of money and manpower, to fully characterize IP for complex SoCs. Even standard IP must fit into non-standard chip configurations with sensitivity to noise, heat and power.

Synopsys’ audio subsystem introduction last month is a case in point. Its decision to bundle its audio chip into a subsystem combined both internal IP and tools with the ARC processor and other IP it acquired from Virage Logic. “The goal was to create a drop-in audio solution that is fully configurable,” said Henk Hamoen, product marketing manager at Synopsys. “We took Virage’s audio codecs, Synopsys’ A-to-D capability, and we added system-level tools virtual prototyping and an assembler.”

That leads to reason two, faster integration. Synopsys’ top rival in this market is Tensilica, which has been developing its own audio subsystems and racking up a long list of codecs to fit into any possible scenario.

“If you were to build the entire thing out of state machines, you would need to lay down 5 million lines of code,” said Steve Roddy, vice president of marketing and business development at Tensilica. “Most successful subsystems have hardware and software and they’re programmable.”

The third reason involves competition. Integrated subsystems raise the barrier of entry for IP vendors to compete in this market. Small IP vendors have a tougher time competing with competitively priced, pre-integrated and verified subsystems—particularly if they can be customized. But there also are logical limits about what gets bundled.

“It may seem natural to an IP provider to put audio and video subsystems together, but no one is asking for it,” said Roddy. “There is a naturally occurring separation. The people who buy DSPs aren’t analog guys so they don’t always understand that side as well the IP vendors do.”

Preparing for stacked die
Subsystems are seen as a great way to speed time to market in complex SoCs, but they’re even more critical in stacked die. Memory subsystems, for example, will likely be entirely separate chips that are included inside the same package.

“The industry in general is looking at more and more integration,” said Vishal Kapoor, vice president of product management at Cadence. “We’ve moved from gate to IP to separate pieces of IP, and in general we’re looking at subsystems—although there is some debate about how to define a subsystem. Connectivity between products appears to benefit from pre-configuration or organizing things together.”

In 2.5D, at least, this is relatively straightforward. Subsystems created in one process may not need to be moved to the most advanced process, particularly when it comes to analog.

“If you have audio DACs (digital-to-analog converters), for example, and they work fine, there’s no need to build new ones” said Tensilica’s Roddy. “If you look at the typical catalog of a mixed signal company, they may have 2,000 to 3,000 parts. A lot of it is older technology, but you can take some of that and for $50,000 turn it into a custom part. You couldn’t do that if you integrated it on a die. This allows you to tweak the analog customer by customer, and to mix and match Legos.”

He said Tensilica has a couple dozen customers in the United States, Europe and Japan working on 2.5D chips, mixing and matching multiple analog parts. That’s can be particularly good news for subsystem vendors. “If you bundle multiple IPs together when it makes natural sense, then you get a multiplier effect for these chips.”

Leave a Reply

(Note: This name will be displayed publicly)