Practical Chip Design

May 6 2008 2:37PM | Permalink | Email this | Comments (2) |
Blog This! using:  Blogger.com | LiveJournal |
Digg This | Slashdot This | add to Del.icio.us

It's a commonplace that a major driver in SoC development is the smart phone. But surprisingly seldom do we see the logical next step: that one of the world's authorities on the direction of SoC design ought to be the most successful player in the smart phone market, Research in Motion, parent of the Blackberry. At a recent conference in Montreal, RiM vice president for silicon Paul Kempf spent some time outlining the future of the SoC from his organization's specialized, but clearly important, perspective.

First, Kempf talked about development cycles. He said that the time for a full handset development—including the silicon—is one year for a completely new platform, and under six months for a derivative product. That leaves not a lot of time for exploration, and none for do-overs.

Then Kempf discussed the physical constraints on a handset. He pointed out that the width and height of a handset--constrained by, among other things, the need to fit in a display and keyboard, and to hold both a microphone and speaker at an appropriate distance from each other--have been mercifully stable for some time. But the thickness of handsets has been shrinking by about 20 percent per year, and this is causing trouble.

Aside from the problem of shrinking the space for batteries, the thinner phones are forcing designers to use metal cases to achieve adequate rigidity, Kempf said. That, in turn, is creating nightmares for RF designers, who must somehow get an adequate amount of energy back and forth between the handset and a basestation antenna, the WiFi hub, the BlueTooth headphone, and whatever, not to mention acquiring and reading a bunch of GPS satellites, all from random handset orientations, while at the same time minimizing the radiation absorbed by nearby brains and other susceptible organs. Now all this must be accomplished from inside a metal can.

As increasing amounts of media processing converge on the handset, the digital side of the design is also growing complex, according to Kempf. He described the video world in particular as a mess of CoDecs and file formats, seasoned by a sharp increase in bandwidth requirements—from about 20 MBytes/s for VGA resolution to about 180 MBytes/s for 1080 p to view video, for instance. Of course these bandwidth requirements directly complicate the problems of the RF designers as well.

But the weight of the challenge falls mostly on the digital designers. Kempf predicted that the current rapid increase in application processor performance would mean handsets would be integrating 1 GHz CPU cores within a few years. That, in turn, has created not just pressure on CPU IP, but what Kempf described as "a constant battle for efficient use of on-chip memory."

The problem looks like this. As the application processor gets faster, it requires more memory bandwidth to achieve high utilization. But the bandwidth of links between SoCs and external memory are not keeping up with the CPU speed increases. (see, for instance, here.) That means that on-chip RAM has to take up the slack between the CPU's bandwidth needs and the off-chip Flash and DRAM's capabilities.

But this same embedded SRAM is being squeezed by the pressure to put additional accelerators, a larger, faster applications processor, and small-signal RF circuitry on the SoC. So the tendency is to cut on-chip memory size, which undermines CPU performance. There is no simple trade-off for this problem, nor any obvious finesse that is not highly use-model-specific.

The RF part of this picture is also a problem, Kempf suggested. Historically, smart phones have kept separate radios on separate dice—often several of them. But with increasing pressure to integrate—both from the shrinking form factor and from cost pressures—smart phones are starting to follow the entry-level phones down the path of integrating the small-signal RF onto the system SoC. "People here are just starting to deal with RF integration," Kempf said. "The idea is to eliminate as many as possible of the multiple substrates that we use now. I think this will end up with all of the connectivity hardware except the cellular radio itself getting integrated."

Thus the problems of the smart phone are growing, wedging the SoC design team between a shrinking form factor and a spiraling list of new functions. No doubt 45 nm will help to some degree, but a bigger transistor budget by itself won't cool a 1 GHZ CPU, triple the effective memory bandwidth of an SoC-based memory hierarchy, or make the small-signal RF stages from a half dozen incompatible radios happily co-operate on a single substrate. There's challenges aplenty ahead.

Related entries in: Convergence | SOC (System on a chip) |