Heard at ARM DevCon: TI, OMAP-3, and the perils of great hardware

As Texas Instruments rolls out its OMAP-3 connected media architecture, we are getting a grandstand view of the interaction between hardware and software development on an advanced platform. The company has lavished a combination of process advances, architectural features and circuit complexity on the architecture in order to provide an excellent cost/performance point for the Web-connected, media-rich applications the designers have in mind. But in the process, they have created the very real possibility of overrunning the capabilities of their own customers. One result has been a closer relationship with operating-systems vendors. Another has been a distinct move from TI’s traditional approach of providing engineering platforms, toward creation of full reference designs.

The sophistication of the new platform begins, according to strategic marketing manager Brian Carlson, with a step beyond state-of-the-market practice in power management. TI has been among the first to employ ARM’s dynamic voltage-frequency scaling (DVFS)—the practice of creating isolated voltage islands within functional blocks, and then adjusting the voltage and clock frequency to the lowest levels consistent with the immediate task load. DVFS was initially supported by ARM on some Cortex CPUs, and a few ARM licensees have extended the technique to other major blocks in their SoCs.

At 45 nm, TI will also use adaptive body-bias—altering the bias voltage on the bulk silicon beneath the channels of transistors to dynamically change their threshold voltages. Unlike DVFS, this requires process modifications as well as significant changes to the design flow.

What both techniques have in common is a dependence on the operating system and application software. This is quite different from clock gating, where you can generally infer the opportunity to turn off the clock to a portion of circuitry from the current state of the block. With DVFS and adaptive body bias, the adjustment process takes a while and itself consumes energy. So you must have foreknowledge of the performance demands that the next few milliseconds will place on a block before you start turning things down.

That means hooks into both the operating system and, in some cases, the application software. To some extent the OS can infer the level of performance needed in a block from the system mode and the information available to the task manager in the kernel. But for really full exploitation of the features, applications need system calls to explicitly request and relinquish a particular level of processing power in a particular engine. That means—if end-products are going to have great energy efficiency—drawing OS providers and application developers deeper into the hardware design.

Nor is power management the only example of this problem. Carlson said that in many cases the real strength of TI’s hardware accelerators goes untapped by algorithms that were developed to be platform-independent, or simply were written before the OMAP-3 specs were available.

This has led TI to a number of steps. One, Carlson said, is much closer cooperation with the handful of OS developers who serve these embedded media markets. This can be a trick, he suggested, as Microsoft is easy to find, and Symbian is easy to find, but there are many paths by which an embedded variant of Linux can end up in a design.

Another major move has been for TI to expand beyond providing drivers for its hardware to providing software for system services, utilities, and in some cases critical tasks at the application level—what the company calls foundation software. Increasingly, TI is finding that rather than customers seeing application code from TI as an intrusion, they see it as major help in an area in which they lacked expertise.

The foundation software has become particularly important, Carlson said, as the goal for the individual applications within a multimedia device has become parity with the best stand-alone devices in the market. Integration no longer justifies compromise, and that means application modules have to be state-of-market, not merely place-holders. And state-of-market software must exploit the hardware features as fully as possible. Which makes it easier all around if TI takes on the responsibility.

The endpoint of this arc is probably a set of complete, production-worthy reference designs from TI. Each would be aimed at a particular class of connected media device, and each would have sufficient knobs on the front so that a system OEM could personalize and differentiate it for their particular view of the market. This TI has done in relatively undifferentiated devices like low-end handsets. But in the high end of the smart phone, mobile internet and Web-connected sub-notebook areas it is a more ambitious goal.