Posted on 17 July 2019

Conserving Energy through Statistical Adaptation

Free Bodo's Power Magazines!




Efficiency is becoming more relevant due to lack of available power

With operating expenses becoming a larger part of equipment purchasing decisions, manufacturers have been continuously looking for ways to reduce the energy consumption of their products. Most look for architectural changes or lower power technologies to incrementally improve the efficiency of their systems while maintaining or improving the performance level. In this article we will examine a concept I will call Statistical Adaptation – a method for improving the overall efficiency by adapting to current conditions.

By Richard Zarr - National Semiconductor


It is widely known that energy prices are increasing as world demand continues to grow. Even in weak economic times, people still need power. With the emergence of cloud computing and the associated increases in infrastructure bandwidth as well as growing capabilities of Personal Mobile Devices (PMD), improving energy efficiency has risen greatly in priority.

In many cases, system architects and engineers have looked to suppliers of semiconductors (and other energy consuming components) to provide lower power alternatives to improve energy consumption. By simply replacing a function with a lower power version, total energy consumption can be reduced. Additionally, by improving the conversion efficiency of power supplies, gains can also be realized. However, there are diminishing returns in attempting to improve efficiencies much past 95%. Cost and size begin to over take the improvements.

There is another method of reducing power in systems through statistical use models that has been applied periodically appearing mostly in portable devices to extend battery life. The idea is quite simple and can be observed in laptop computers… as the user stops typing, the processor is slowed down and if enough time goes by with inactivity (determined by settings), back lights are reduced and hard drives shut down. The system “adapts” to the users behavior and statistically the laptop will run longer than if these features were turned off.

Continuous Adaptation

The previous example illustrates gross adjustments in system power consumption accomplished mostly by either slowing clocks or powering off subsystems. There are finer levels of adaptation that can provide even higher levels of power savings. This is referred to as continuous adaptation which uses extremely fine levels of adjustment and closed loop monitoring to get extremely close to the lowest possible power levels without failure.

An example of such a system could be an LED back light controller for a DVD player. In this scenario, the back light cannot be simply turned off if the user stops typing for a few minutes. The unit is used continuously while watching a video, however the back light can adapt to the viewing environment. As ambient light levels change, the backlight can be adjusted to compensate. Users that always watch their DVD player in very bright light will not see an advantage to this method, but statistically users that watch videos in average light will see improved run times. Figure 1 shows a Gaussian distribution approximating the ambient lighting model for a DVD player. Other distributions could be used as well depending on other usage models, but we’ll use the classic “bell curve” and associated error functions. At any point in time most users will be between -3ó and 1ó which represents around 84% of the population (a majority). This also represents a light intensity range from almost complete darkness to moderately bright – where most of us would watch a DVD.

Gaussian distribution of DVD light levels

To understand the average amount of energy saved based on this distribution, we’ll need to calculate the area intersected by the linear function of brightness (dark to full sun) for all users and distribution function from figure 1. We should assume the lowest light level would be 20% at -3ó and 100% at +3ó. This is illustrated in figure 2 and equation 1. The calculations are shown in figure 3.

Relative Power over light levels

Equation 1 - Backlight Level


As you can see from the calculations, only 42% of the back-light energy (relative to “full on”) would be consumed by the 84% of the population that watch their DVD player in dark to moderate lighting conditions. That is if a continuous method of back-lighting control is implemented that senses ambient light. Even if we include the remainder of the distribution, the energy consumed only rises to around 60% (assuming a Gaussian distribution). Not everyone will see such an improvement, but over a large population these savings can be observed.

Adaptive Power

The same applies to systems that vary in other parameters such as temperature, process or aging as found in CMOS semiconductor devices. The power consumed in modern digital devices can be very large (see equation 2). The larger component of the power dissipation is the dynamic component (V2 term) due to the exponent. However, in modern small geometry processes, the gate and sub-threshold leakage can be quite significant. Both components are dependant on supply voltage which provides a means of control over the power dissipated if a system could monitor the process performance.

Equation 2 – CMOS Power Dissipation

Figure 4 shows a typical distribution of performance and the required Vdd supply voltage. Engineers must always use the worse case parameters to guarantee a system will operate over all conditions. In this case, the supply voltage would need to be 1.2 volts to make sure all chips from this process will operate at frequency over temperature and aging. However, if an embedded performance monitor was added to the design and a closed loop system devised to control the Vdd on each device, then the supply voltage could be adapted to provide the exact amount of energy required to operate – not the worse case. National Semiconductor’s Adaptive Voltage Scaling is an example of this technology.

CMOS Process Variation

Applying a similar analysis on the example Gaussian curve shown in figure 4 yields a theoretical savings of over of 30% (dynamic only) in a system with many large digital devices or many systems with one large device (See equation 3). The lower energy consumption is accomplished simply by controlling the supply voltage to match the energy requirements of each individual device while maintaining timing closure. It does not include additional techniques such as frequency scaling which will improve the energy savings even further.

Percentage dynamic power saved over distribution (Vdd = 0.8V to 1.2V)


If all systems were created identically and used in the exact same manner every time, the worse case condition would always be the best case condition. However, in our world of ever drifting environments, processes and user tendencies it is best to create systems that continuously adapt their power consumption to the current system state. Energy efficiency is becoming more relevant due to user preference, regulation and simply a lack of available power such as in remote installations. Adaptive power management can greatly improve system energy consumption and reduce operating expenses in large scale installations such as network infrastructure and cloud computing as well as in personal mobile devices.



VN:F [1.9.17_1161]
Rating: 0.0/6 (0 votes cast)

This post was written by:

- who has written 791 posts on PowerGuru - Power Electronics Information Portal.

Contact the author

Leave a Response

You must be logged in to post a comment.