Talking Heads

The use of more third-party IP inside SoCs coupled with problems encountered at advanced process nodes is turning up some interesting challenges—and pointing the industry in some interesting directions.

It’s a well-known fact that third-party IP isn’t always used as it was intended. Even internally developed IP isn’t always used as prescribed. It’s not unusual for chip developers to turn up the performance, turn down the power, or to introduce unanticipated noise around a piece of IP. This is all normal and part of designing a chip. Engineering managers grouse about the trouble with integrating IP, software and sometimes even the package, but in the end—at least until recently—it could all be fixed with some last minute tweaks—even after tapeout.

Companies are finding that isn’t always the case at 32/28nm, and at 22/20nm there will be even less leeway in repairing the damage. Moreover, there isn’t the kind of margin needed to provide extra padding for these kinds of potential problems. At advanced nodes, guard-banding boosts power beyond acceptable limits or reduces the overall performance of an SoC—or both.

What’s required at advanced nodes is a lot of give and take from all parts of the design team throughout the process. For companies that have design teams scattered around the globe, this presents a new problem. The days of labor arbitrage are ending. The system has to be tightly integrated, and parts need to be adjusted as the chip progresses at each stage by people who can address those issues in real time. That may mean tweaking everything from the RTL to the package, the software, and potentially even the board. For companies that have put teams in various time zones to keep development moving round the clock, this isn’t always possible.

At future nodes, it may even affect the entire ecosystem. For a disaggregated model to work with really complex designs the ecosystem has to function like a well-oiled machine. Scattering it across different time zones with language barriers doesn’t help. Scattering it across different disciplines, with their own language problems, doesn’t help either—particularly when those disciplines aren’t under the same engineering management. And while those problems may seem interesting on a passing note, they will become rather glaring as we move to advanced nodes and 3D stacking.

There has been a lot of talk about the advantages of IDMs and vertically integrated companies, but the real advantage probably has less to do with ownership than structured and open communication channels. The companies that figure out how to do that best will be the ones that will emerge as the next leaders in this industry. It appears that communication has suddenly become very important to system-level design. Unfortunately not everyone does it well, and fewer people do it well all the time and across multiple disciplines. It’s probably a good time to learn.

–Ed Sperling