Power Control Moving into Hardware

Sonics have been building networks-on-chips (NoCs) for a long time and have amassed a rich patent portfolio. So being granted a new one isn’t usually deemed press-release-worthy. However, their latest patent on power management is pretty significant. It is patent 8,601,288 titled “Intelligent Power Controller”.

Historically SoCs have had a manageable number of blocks and a limited number of power domains that could be powered down. The responsibility for actually powering them down and bringing them back up again was the responsibility of the embedded software. Even then, chip designers would complain that the software developers didn’t power down blocks aggressively enough and the embedded programmers would complain that they didn’t understand all the implications of setting bits in the power registers, how much power would be saved, how long before they needed to use the block must they power it up and so on. Just as with the move to multicore processors, the hardware people pushed their problem up into the software world and assumed that the software people would be able to solve it without much difficulty.

Now SoCs have a couple of hundred blocks and the software architecture has also got very complex. Think of iOS or Android, with a portfolio of apps loaded onto the hardware after it shipped, compared to, say, the software in a 5-year old digital camera with a lightweight real-time operating system (RTOS) and a single application. How can a part of the software even know whether it can power down the LTE modem since other parts of the software may be using it to access the internet. When a phone was either making a call or not making a call, life was a lot simpler.

NoCs offer an opportunity to move the responsibility for powering blocks up or down into the hardware. The Noc knows two useful things: whether a block is receiving or transmitting anything, and whether another block is trying to send a block something. Blocks can automatically be powered down when they have been idle for too long, and when another block tries to communicate with a powered-down block then the NoC can buffer the message, power up the block, and then deliver it. Of course this needs to be done with some input from the architect. If it takes too long to power up a block then this automatic approach may not work. And some blocks might want to be powered up whether anyone is communicating with them or not if they run some continuous monitoring process, for example.

All of this is complementary to other power saving approaches such as clock-gating, multi-Vt cells and so on. Power is too important not to address it at every level from low-power processes up to choice of algorithms that are power-sensitive. But one of the biggest possible power savers is to do what your parents told you: turn the lights off when you are not using them.

Sonics press release here. Full patent here.