16/11/2018
COMPUTER FUNDAMENTAL - Power
Management
What is Power Management
Power Management is a feature of some
electrical appliances, especially copiers,
computers, GPUs and computer peripherals such
as monitors and printers that turns off the power
or switches the system to a low-power state
when inactive. In computing this is known as PC
power management and is built around a
standard called ACPI. This supersedes APM. All
recent (consumer) computers have ACPI
support.
Motivations
PC power management for computer systems
is desired for many reasons, particularly:
Reduce overall energy consumption
Prolong battery life for portable and
embedded systems
Reduce cooling requirements
Reduce noise
Reduce operating costs for energy and
cooling
Lower power consumption also means lower heat
dissipation, which increases system stability, and
less energy use, which saves money and reduces
the impact on the environment.
Processor level techniques
• The power management for microprocessors
can be done over the whole processor,[2] or in
specific components, such as cache memory[3]
and main memory.[4]
With dynamic voltage scaling and dynamic
frequency scaling, the CPU core voltage, clock
rate, or both, can be altered to decrease power
consumption at the price of potentially lower
performance. This is sometimes done in real time
to optimize the power-performance tradeoff.
Examples:
AMD Cool'n'Quiet
AMD PowerNow! [5]
IBM EnergyScale [6]
Intel SpeedStep
Transmeta LongRun and LongRun2
VIA LongHaul (PowerSaver)
Additionally, processors can selectively power off
internal circuitry (power gating). For example
• Newer Intel Core processors support ultra-fine
power control over the functional units within the
processors.
• AMD CoolCore technology get more efficient
performance by dynamically activating or turning
off parts of the processor.[7]
Intel VRT technology split the chip into a 3.3V I/
O section and a 2.9V core section. The lower
core voltage reduces power consumption.
Power Management in GPUs
Graphics processing unit (GPUs) are used
together with a CPU to accelerate computing in
variety of domains revolving around scientific,
analytics, engineering, consumer and enterprise
applications.[8] All of this do come with some
drawbacks, the high computing capability of
GPUs comes at the cost of high power
dissipation. A lot of research has been done over
the power dissipation issue of GPUs and a lot of
different techniques have been proposed to
address this issue. Dynamic voltage scaling/
Dynamic frequency scaling(DVFS) and clock
gating are two commonly used techniques for
reducing dynamic power in GPUs.
DVFS Techniques
Experiments show that conventional processor
DVFS policy can achieve power reduction of
embedded GPUs with reasonable performance
degradation.[9] New directions for designing
effective DVFS schedulers for heterogeneous
systems are also being explored.[10] A
heterogeneous CPU-GPU architecture, GreenGPU
[11] is presented which employs DVFS in a
synchronized way, both for GPU and CPU.
GreenGPU is implemented using the CUDA
framework on a real physical testbed with Nvidia
GeForce GPUs and AMD Phenom II CPUs.
Experimentally it is shown that the GreenGPU
achieves 21.04% average energy saving and
outperforms several well-designed baselines. For
the mainstream GPUs which are extensively
used in all kinds of commercial and personal
applications several DVFS techniques exist and
are built into the GPUs alone, AMD PowerTune
and AMD ZeroCore Power are the two dynamic
frequency scaling technologies for AMD graphic
cards. Practical tests showed that reclocking a
Geforce GTX 480 can achieve a 28% lower power
consumption while only decreasing performance
by 1% for a given task
Power Gating Techniques
A lot of research has been done on the dynamic
power reduction with the use of DVFS
techniques. However, as technology continues to
shrink, leakage power will become a dominant
factor. Power gating is a commonly used circuit
technique to remove leakage by turning off the
supply voltage of unused circuits. Power gating
incurs energy overhead; therefore, unused
circuits need to remain idle long enough to
compensate this overheads. A novel micro-
architectural technique[14] for run-time power-
gating caches of GPUs saves leakage energy.
Based on experiments on 16 different GPU
workloads, the average energy savings achieved
by the proposed technique is 54%. Shaders are
the most power hungry component of a GPU, a
predictive shader shut down power gating
technique achieves up to 46% leakage reduction
on shader processors. The Predictive Shader
Shutdown technique exploits workload variation
across frames to eliminate leakage in shader
clusters. Another technique called Deferred
Geometry Pipeline seeks to minimize leakage in
fixed-function geometry units by utilizing an
imbalance between geometry and fragment
computation across batches which removes up
to 57% of the leakage in the fixed-function
geometry units. A simple time-out power gating
method can be applied to non-shader ex*****on
units which eliminates 83.3% of the leakage in
non-shader ex*****on units on average. All the
three techniques stated above incur negligible
performance degradation, less than 1%.
