The entire semiconductor industry is worried about power consumption, but almost nobody is looking at the right solutions.
These days it seems that the entire semiconductor industry is worried about power consumption, and is doing everything it can to address the issue. Multicore architectures have been the most prominent solution. These include everything from dual-core Pentiums to the new Ambric Am2045, which contains a stunning 360 processors inside a single chip.
Chip vendors are also starting to use "holistic design" approaches that incorporate strained silicon, low-k dielectric insulators, aggressive clock gating, and many other power-saving technologies. They are even looking at extreme techniques such as using error correction to allow for "sloppy" chip designs.
The problem is that most of these approaches improve power consumption only under the right conditions. For example, multicore chips are a big help for certain applications—particularly those with multiple independent channels—but it is not clear that they provide benefits across the broad spectrum of DSP applications. In fact, some in the industry believe that multicore architectures can increase power consumption.
The bigger problem, though, is that almost all of the power-saving efforts have been focused on processors. This is a problem because processors are often a minor player in the overall power consumption of a device. In PCs, for example, the power supply eats up an incredible 50 percent of the power! With these kinds of losses, you could drop the processor power to zero, and still not see a big improvement at the system level. For another data point, consider this table (which is taken from this article):
In this example, reducing the DSP power by 50% only reduces the board-level power by 25%. Even more noteworthy, observe that most of the board-level savings came from moving code on-chip, not from optimization of the DSP power.
The point is that saving power requires big-picture thinking. Effective low-power design involves everything from selecting the right A/D converters to minimizing the trace lengths on your PCBs. Getting all of this right is a daunting challenge. Clearly, we need help.
Therefore, I issue this call to arms to the silicon vendors: Spend less time squeezing microwatts out of your processors, and spend more time helping us get our system-level design right. Give us reference designs, training, consultation—whatever it takes.
And now a call to arms for engineers: what resources do you need to tackle low-power design? Visit the forums and let me know.