Will kHz be king for IoT applications?

Far from simply being a matter of including a web interface to your device, IoT is posing some more fundamental technological questions. Mike Muller (pictured), ARM's chief technology officer, addressed some of them at a recent lunch in London. One of the basis questions is what kind of processor might IoT devices need? "Does IoT need 32bit processors? Yes," he said. "The incremental cost is small from the hardware perspective, but IoT is being driven by software complexity. It doesn't take much to 'break' 8 and 16bit devices and users will see it's easier on 32bit." He also contended that dealing with the protocol stacks would be easier on 32bit parts, rather than segmented on an 8bit device. "It's not easy to run ZigBee, for example, on an 8bit micro," he commented. But another looming issue for IoT is power consumption. Muller said IoT applications will need very low power processors and implied these devices are beyond what ARM currently has on its road map. One approach which is being pursued by a number of researchers is that of near threshold and sub threshold switching. The threshold is the point where the transistor switches and starts to conduct. Depending on how the transistor is designed, this can be as little as 0.3V. The concept says that if the supply voltage is maintained at or near the threshold, switching will take place but more slowly. By adopting near threshold switching, designers trade speed, but gain power consumption benefits. "We've built a lot of sub and near threshold designs," Muller noted. "Theoretically, it's the right place to be." Sub threshold processors may run at kHz rates, rather than GHz. While the logic behind this approach can be appreciated, the realities of building such devices commercially are a challenge. "Foundries are more confident with near threshold designs," he said. "Near threshold is easier because it performs closer to the Spice models, rather than what you think the process might be." Even when running at low supply voltages, Muller believes these processors will still need additional measures to conserve power, including clock gating and power gating. "There are implementation choices to be made," he accepted, "When and how do you do power gating and clock gating?" he asked, adding that new techniques – such as clock gating within a clock cycle, rather than between cycles – might be needed. Such an approach might well find application in energy harvesting and energy scavenging applications, rather than in battery power designs. "The gating schemes will be different," Muller believed. "Should the processor run quickly, then stop or should it simply run slowly or should it operate in response to the amount of energy available?" Another approach being considered by ARM is drowsy logic. Muller explained the concept. "With drowsy logic, when you turn the power off, the voltage rails start to decay slowly. If you can turn the power back on before the voltage drops too low, you have state retention. If you can stop at the right place in the process, you can let the voltage rails decay. The flip/flop still has retention and the amount of power needed to restore the output is small." As to progress with these techniques, Muller was non-committal, but said: "We can do these things with a Cortex-M0 core; it's all about the implementation."