1. Power Reduction Techniques For Microprocessor Systems
Presented by William Lane
Material and graphics presented here are from:
Power Reduction Techniques For Microprocessor Systems
Vasanth Venkatachalam and Michael Franz
ACM Computing Surveys, Vol. 37, No3, September 2005, pp

2. “Power consumption is a major factor that limits the performance of computer systems.”
Some of this power is used in a very small area which would mean a high power density. As you can see from the chart that modern processor are reaching power density level that are comparable to that found in a nuclear reactor. As chips have gotten more powerful, as to has their heat sinks and fans. When chips run hotter then generally their reliability is lower and their health can be impaired.
On your home computer you probably would trade a small amount of power consumption to be done slightly quicker. When buying a processor for your home desktop computer what do you consider? Price and Speed. Notice that you don’t care about power it uses.
Now how about your cell phone? When you buy a cell phone would you rather have a phone with a shorter life but some extra graphics processing or a longer battery life without many bells and whistles?
Most of the power a computer consumes ends up as heat.
The above diagram shows the power density of a few microprocessors.
In some cases heat generation and power consumption may not be an issue.
For example: Home computer used for gaming vs. Cell phone

3. Forms of power consumption:
Dynamic
Switched capacitance
Short circuit current
Static
Leakage
Power: rate at which computer consumes electricity, or dissipates it as heat
Energy: total electricity the computer uses
Dynamic power loss: refers to circuit activity such as changes to a register
Static power loss refers to energy that is lost while not doing any useful work. Like a capacitor that slowly loses charge over time.
This graph was from 2005 so the levels above 65% are speculative.
Static power loss seems to be where most of the total power is ending up.

4. Dynamic Power Loss Switched capacitance: Short Circuit Current:
85-90% of dynamic power loss
Energy used in charging or dissipating capacitors at outputs of circuits
Short Circuit Current:
10-15% of dynamic power loss
Energy used for pairs of transistors that switch current with opposite polarity cause a momentary short circuit

5. Equation for Switched Capacitance power loss
P = aCV2f
P = power loss a = activity factor
V = supply voltage f = clock frequency
C = capacitance
As you can see the area you would want to focus on first is the supply voltage

6. Methods of reducing Switched Capacitance power loss
Decreasing the capacitance
This can worsen the performance.
Decreasing activity level
This can be difficult because generally we try to fit more on to computer chips. Ex. Clock Gating
Reduce clock frequency
Reduces the performance and may not end up reducing the energy used.
Decrease supply voltage
Lower supply voltage can increase time delay of the circuit which would then need a lower clock frequency.
There are a few methods used to decrease activity level. This includes clock gating. This is where some part of the chip is able to be turned off and even the clock signal is not transmitted to it. This method is used on the P4 and PM. A disadvantage is there can be some lag time involved with starting this sections of the chip back up. To do this one may have to add another stage to the EXE part of the pipeline to turn on the part of the processor needed.
As you saw in the previous slide a reduction in supply voltage has the largest effect on power loss.

7. Dynamic Voltage Scaling
A combination of adjusting the supply voltage and clock frequency are used to increase or decrease the power of the processor. Adjusting these two together gives the best results.
Many modern processors use such a method to control the speed of the processor.
DVS can be used for consuming less power while also keeping the temperature of the chip cooler.
DVS allows for increased computer processing when needed and then once done it can then run the processor slower. There are many type of algorithms used for detection of when to run at a higher clock rate and when not to. Some processor continually watch the delay in the circuit to decide whether the clock cycle should be increased or decreased. Some processor have an temperature sensor built into the chip. When the chip gets to war it then switches to a cooler setting that hinders system performance slightly.
One place DVS can be taken advantage of is if there is a stall in the processor, like during a memory access. If the processor is at idle with not task to do then a good idea would be to reduce the power consumption while waiting. This only would really work when there is a minimal penalty for changing clock cycle frequency.

8. Static Power Loss Reduce supply voltage Reduce the size of the circuit
Uses less power
Reduce the size of the circuit
Clock gating
Technology Mapping (tool for circuit design)
Cool computer
Recent study showed that cooling the processor by 50 degrees reduced leakage by up to a factor of 5
Change threshold voltage of transistors
Increasing the threshold voltage lowers the amount of leakage through the transistor, but it then takes more voltage to switch the transistor. Ex: Adaptive Body Biasing
This is the most disappointing of power loss. It’s similar to your DVD player being turned off but consuming power because it is plugged in.
As you can see some of these areas are the same as that of Dynamic Power Loss.
Technology Mapping: place transistor that change often closer to the end of the circuit. This reduces the number of capacitors that must switch. This problem is an NP-Hard problem.
Cooling chip: electricity has less resistance in lower temperatures. Cooling helps with the overall health of the processor. Cooling must be uniform across chip.
Threshold voltage: one idea is to on path that are frequently used to have a lower threshold and on paths that are rarely used to have a higher threshold. This can be difficult to decide as one designs the circuit and a newer method using transistors that can adjust the voltage threshold is called ABB.

9. Further studies in power consumption
Buses
Reduce the amount of switching in a bus. Invert signal to decrease number of bits that change.
11111 to bits change
11111 to bit change
Low Swing Bus
Use lower voltage for the bus
Bus segmentation
Shut down parts of bus not used
Bus inversion has a cost an extra bit to transmit and some additional hardware invert signal upon receiving it
LSB: lower voltage from normal 5V down to 0.3 and then back to 5V a receiving end.
Bus segmentation: Make the bus shorter. Since the entire bus must be brought up to the correct voltage this makes the bus smaller and uses less power.
Some research is also going on with replacing buses with a network.

10. Further studies in power consumption
Memory
Each memory access costs energy
Decrease number of memory access
Decrease amount of energy needed to access memory
Scratch pad memory
Used mostly in embedded application; programmer defines what data to store in the cache.
Instruction level cache
Stores the next few instructions only. Any memory references are removed. Located between L1 cache and processor
Adaptive Cache
Turns off part off cache to save energy. Can be used with drowsy cache to save data in cache when cache is turned off.
Adaptive cache: as workload increases on processor more memory is turned on. This is done by watching the number of cache misses and when it gets above a certain level more cache is turned on.
Drowsy cache: Some parts of cache go into a low power mode where just enough power is given as to not lose the data. IF the cache is to be accessed the cache must be brought back up to normal power to access it. This adds time until getting data from sleeping cache.

11. A few open questions
Is it more energy efficient to run a program using a more powerful processor which runs for less time, or is it better to run a processor that is less powerful but finishes in more time? Which processor uses more energy?
The speedup of a program is not linear to the speedup of the processor.
Is it better to use multiple slow processors or one fast processor? Which uses more energy?
For a laptop computer would you pick a processor that would burn through your battery quickly but gives you more computing power?
Further more with a more powerful processor in your laptop you also have to worry about heat dissipation. Most laptop are designed small and thin which makes it difficult to put large heat sinks and fans in them.
Another example of this is, Is it faster to move a group of people form Point A to point B with a sports car or a school bus. Providing that the bus can carry 30 times as many people as your sports car but the sports car and go on average 6 times as fast as the school bus. Which is better?
First question may be how many people are we talking about? Then there are many other considerations as well.

12. In conclusion
The energy efficiency of a processor needs to be weighed against the application of the processor.
Most often it is a choice between a more powerful processor and a more energy efficient one.