Designing a low power CPU

First up, please remember that I am a software guy, not a chip designer. Please read this posting with that in mind.

I was chatting with someone at a conference recently and they planted an idea in my brain. I am not sure what to do with the idea, so I thought that I would exercise it here. The topic was the consideration of power consumption in embedded systems – a subject I have written about before: here, here and here for example. But this was a new angle …

Until recently, low power design was simply seen as a hardware issue – just a matter of using the right electronic design to minimize power. Of late, it has become the province of software engineers, but for this to be possible, the hardware design needs to accommodate the needs of software by providing:

• a means to adjust voltage and frequency, as required to optimize power/performance
• the ability to turn off unused peripherals
• access to low power modes – hibernate, standby etc.

This is fine, but the idea proposed was that the CPU could be optimized to reduce power by careful design of its instruction set – or, to be more precise, the binary encoding of instructions.

Traditionally, the binary codes of CPUs’ instruction sets have been designed using a number of bit fields, which designate the type of instruction, mode of operation, registers in use, etc. This seems logical and makes the binary code easier to analyze and understand and somewhat simplifies the instruction decoding. It is this last part which is interesting, as the simplification is not drastic. If arbitrary numbers were assigned to each unique instruction, decoding would still be reasonably straightforward. There has been no incentive to take this approach – until now.

In an electronic circuit, a key factor that affects power consumption is a logic transition – 0 to 1 or vice versa. So, in theory, minimizing the number of transitions would reduce power consumption. When a CPU executes a sequence of instructions, there are logic transitions associated with the difference in the binary code of one instruction compared with the next. If the binary instruction set could be designed to minimize transitions, power consumption would be reduced.

This brings us back to software, because the code determines the sequence of instructions. All that needs to be done is the analysis of a large body of typical embedded code, looking for common sequences of two or more instructions. These instructions could then be assigned values that differ in fewer bits [preferably just one].

I have no idea whether this idea is original or whether it has been deployed. Any input by comment or email would be appreciated.