Intro to Computer Systems

Chapter 9: Power Considerations

Sources of Inefficiency

Nothing is constructed perfectly - there are always inefficiencies in any kind of mechanical or electrical design. In computers, these inefficiencies manifest themselves as power draw outside of useful work.

Inefficiency can be through three means:

Intrinsic and Operational Inefficiencies

For the most part, not much can be done about intrinsic or operational inefficiencies; they are a consequence of using a certain technology to build a device. Through careful design the negative side effects can be reduced, however they can never be completely solved.

Semiconductors: Transistor Gate Leakage

Silicon transistors use power when switching as part of their normal workload, known as their dynamic power. However, they also use power from simply being powered on and idle; this is known as static power.

In a perfect world, the technology used to build modern integrated circuits (complementary metal-oxide-semiconductor, or CMOS) would not use any power if not being switched; yet desktop CPUs and GPUs still consume a dozen (or more) watts of power even when completely idle.

The cause of static power usage in processors is a phenomenon known as gate leakage. This is due to the microscopic construction of the transistors on the silicon substrate - the surfaces of the transistor mean are not perfectly insulated from each other, and some power 'leaks' from one part of the transistor to another.

Unless semiconductor fabrication processes improve to concentrate, this leakage current gets worse when transistors are shrunk.

Intel had serious problems in 2004, when they shrunk their CPUs to the "90nm" (90 nanometer) process - their fabrication technology at the time could not deal with the increased gate leakage at this size, causing a rapid increase in static power dissipation.

Progress in Intel's transistor fabrication process.

It was an issue that was only solved with improved transistor construction technology and materials.

Power Electronics: Analog Devices

Power supplies are analog equipment; they do not work on binary data, but continuous signals - the signal in this case being the supply power. Furthermore, modern computers are complex devices: not only must the power source be carefully regulated, but different parts of the system require different voltages, and each have their own power requirements.

The internals of a computer power supply unit.
The internals of a computer power supply unit.

The power supply is a complicated power transformer, that supplies several different voltage circuits, known as rails, to the computer. Some rails have the same voltage; they are duplicated mainly to ensure that no one part of the power supply has to bear the entire power demand alone.

Modern power supplies are based on a switch mode power supply design. In an ideal world, this design would not lose any power when idle, but when built in the real world with large transformers, transistors and discrete electronic components like capacitors and inductors, efficiency is lost due to the internal resistance of these components.

Hard Disks: Friction and Noise

Hard disks are the best example of mechanical components in a computer system. Like other mechanical components, there are friction losses involved with motors and moving parts. The motors and bearings of a hard disk are of a very high quality; however again, it is not a perfect world and the friction in these parts is the cause of much heat and noise.

Although the platter area of the hard disk is kept airtight, it is not a vacuum: further friction (mainly resulting in noise) is generated by the sound of the disk platters spinning very fast in air - a 'whoosh'ing sound.

LCD Monitors: Backlighting

An LCD monitor relies on light passing through the liquid-crystal element, which either allows light to pass through, or blocks it. Thus, the entire back side of the LCD plane must have a light source shining on it.

The backlight of an LCD panel is what produces this light, and is the primary source of power consumption. There are two technologies used in LCD backlighting:

CCFLs have disadvantages, such as their fragility, and reliance on the inverter which increases complexity and decreases long-term reliability. LED backlighting is a newer technology that is undergoing rapid development (as opposed to the largely mature CCFL tubes), and outpaces the older technology in power efficiency, longevity, and also image quality.

Design Inefficiency

Some power consumption inefficiencies are purely due to poor component design and selection. These kinds of power budget blowouts are entirely preventable, either through components of higher quality, or those with a performance rating more in line with the system as a whole.

Lack of Integration

There are many reasons for putting many peripheral interfaces onto highly integrated chips; one major reason is that it reduces the chip count, and thus the complexity of the circuit board. This has an additional advantage: it is no longer necessary to have the required 'glue logic' to appropriately run a number of chips, cutting down on power draw.

For example, consider the case of a chipset southbridge with integrated Serial ATA controllers, to an older southbridge that would rely on a PCI expansion card to provide Serial ATA connectivity. There is additional power draw not just in having a discrete Serial ATA controller chip on the PCI card, but also the support circuitry for that controller chip, interface circuitry to connect to the PCI bus, the additional workload on the PCI bus (and controller), ... this is far less power efficient than adding the functionality onto the southbridge directly.

Overspecified Components

Hardware that is more powerful than necessary can actually lead to sub-optimal performance, when it comes to power consumption.

Power Supplies

The classic example is of computer power supplies.

Sample power supply efficiency curves.
Sample power supply efficiency curves.

The efficiency of power supplies varies with the load placed on them: they have a 'sweet spot' where they are most efficient, usually between 20% and 90% of maximum rated load. At very low loads, less than 20% of their rated maximum power output, they are actually very inefficient - only 60% or so of the power going into the power supply turns into properly regulated output power. The rest turns into heat.

Thus, matching a modest computer system (which may only require 70W at idle), with a very highly specced power supply - 800W maximum power or greater - is a very poor choice. The power supply will never be able to work in its efficiency sweet spot.

AnandTech has an article on debunking power supply myths.

Anything where a high performance component has a high idle power output is a candidate for optimisation - for example, high-end processors and graphics cards will have very high idle power consumption. If the bleeding-edge performance of these components is not required, then the system will be expensive to own and operate, with no positive effect.