The consumption of power in CPU CMOS circuitry may at first seem like a simple matter, but trade-offs with cooling and processing performance make power optimization trickier than simple minimization. For example, for portable or hand-held systems, battery life and limitations of size and mass make active cooling difficult and favor power management and passive cooling solutions. Smaller features, such as those found in chips based on 90-nanometer technology, require lower voltage and can operate at high clock rates, but at the same time lead to quadratic increases in the number of transistors.
By comparison, in a large datacenter, higher CPU voltages can provide faster signal propagation and switching times, but result in higher power consumption and more heat generation. Active cooling can reduce junction temperatures and increase supply voltage, providing faster switching and better performance. For datacenters, the density of CPUs is often limited by heat generation, which can greatly reduce reliability if not controlled. System architects should be clear on the relative priorities of power consumption, heat generation, and performance. Understanding power consumption and the heat generated by it is a good place to start; Equation 1 helps quantify these issues.
Equation 1. Power consumption estimation for digital CMOS circuits
Paverage = Pswitching + Pshort-circuit + Pleakage
In Equation 1:
- Pleakage is the power loss based upon threshold
- Sprobability is the probability that gates will
switch, or a fraction of gate switches on average.
- CL is load capacitance.
- Ishort is short circuit current.
- fclk is the CPU clock frequency.
Based on the power estimate in Equation 1, methods to reduce power consumption include:
- Reducing voltage, since the Vsupply2 term is a quadratic multiplier. However, keep in mind that reducing voltage will increase leakage.
- Reducing clock frequency, slowing down the CPU pipeline and instruction throughput if possible based upon application performance requirements.
- Shutting down the CPU when not in use, saving state to NVRAM.
- Using gated clocking and asynchronous logic where possible to minimize clock distribution.
Many embedded systems architects are more concerned with keeping power consumption and heat generation low than they are with improving performance. Obvious examples of systems designed with these goals in mind include cell phones, PDAs, and laptops, where battery life is a major limiting factor for users. Processors targeted for these systems, such as the PowerPC® 405LP (see Resources), operate at low voltages with dynamic voltage scaling (DVS) for clock rate control to provide high MIPS/milliwatt performance. The MIPS/milliwatt ratio is more important than MIPS alone for battery-powered systems, which also need compact and simple passive cooling.
In contrast, large-scale telecommunications and datacenter embedded processors are not so limited in terms of available power, but still must handle cooling for reliability, and thus might incorporate sophisticated active cooling to improve performance.
Cooling electronics is often a matter of transferring heat to maintain systems at room temperature. (Typically, you should not cool systems to below room temperature; doing so can lead to problems like condensation from the ambient environment.) Cooler systems perform better and are more reliable. Many CPUs would reach temperatures higher than 100 degrees Celsius without active or passive cooling, greatly reducing the system's lifetime, reliability, and performance.
For portable systems, these temperatures would also lead to usability issues. According to the zeroth law of thermodynamics (see Resources), two objects with the same temperature are in equilibrium and no heat flow will occur between them. Devices like CPUs, using significant power (up to 100 watts, in some cases), do work and create waste heat, raising the device temperature; the temperature can only be lowered by removing heat through conduction, convection, or radiation. These three heat transfer modes can restore equilibrium between the ambient room temperature and the devices.
Systems like those designed for the z990 provide active refrigeration and cooling methods, and can actually hold the CPU devices to room temperature. With passive cooling methods, the device temperature is determined by the efficiency of heat transfer out of the device to the ambient surroundings. Equilibrium is achieved when the heat generation rate matches the heat transfer rate.
Equation 2. Thermal conduction
Heat-Flux = -k (dT / dx)
In Equation 2:
- k is the thermal conductivity of a material in watts/meter x Kelvin (for example, 177 for an aluminum alloy).
- dT/dx is the thermal gradient.
- Heat-Flux is watts/meter2 and the heat transfer is determined by the area of the conductor.
Conduction is a simple passive method, but can require large areas and high thermal conductivity to match the power generation of devices. For example, a 100-watt device operating at 50 degrees Celsius would require a copper radiating area 22.6 centimeters2 in area and 10 centimeters thick to reach room temperature. The heat rate is the product of heat flux and the area of the conducting surface. The thickness of the conducting material has a resistivity (insulating property) keeping heat from flowing. So, ideal conducting designs maximize area and minimize thickness of materials. This is why CPU radiators cover a large surface with thin finned radiators.
Equation 3. Thermal convection
Heat-Flux = h (Tsurface - Tambient)
In Equation 3:
- h is the convection heat transfer coefficient in watts/(meter2 x Kelvin) (typically 25 to 250 for forced air).
- Heat-Flux is watts/meter2, and the heat transfer is determined by the thermal boundary layer.
The determination of h for convection is not simple. It requires significant fluid mechanics modeling for a specific surface and boundary layer. Of course, the operation of a fan also requires more power consumption. System architects generally choose convection cooling due to its low cost and simplicity rather than for performance or reliability reasons.
Equation 4. Thermal radiation
Heat-Flux = Emissivity x Boltzmann-constant x (Tsurface4 - Tambient4)
In Equation 4, Heat-Flux is watts/meter2 and the heat transfer is determined by the emissivity of the radiating surface.
Of the three methods of heat transfer, radiation is the simplest: it merely requires a large area with good emissivity properties to transfer a large number of watts to the surroundings. The larger the area of the radiating surface, the more total heat that can be radiated from a hot CPU to the cooler ambient environment.
Conduction and radiation can be implemented with a completely passive heat transfer system, whereas convection is an active approach that requires work input. Most often, a combination of conduction and radiation are used for embedded devices. A fan for convective cooling might be added.
The z990 uses active refrigeration to cool MCMs (multi-chip modules) down to 25 degrees Celsius for two reasons: high performance and high reliability. Power consumption is a lesser concern for this big iron machine. Refrigeration systems are complex, costly, and more likely to need maintenance. The z990 refrigeration design minimizes cost and complexity, but the designers, realizing that the refrigeration system could potentially fail, made the z990 fail-safe. When the temperature of the MCMs climbs due to a refrigeration system malfunction, the CPU core clock rates decrease so as to reduce power consumption and corresponding heat generation. At the lower clock rates (cycle steering), a convective fan cooling system is used as a backup so that the system can continue to operate, albeit in a performance-degraded mode.
Most embedded systems can't afford the cost and complexity of refrigeration; nevertheless, the hybrid cooling approach, the fail-safe design, and the cycle steering are all concepts that can still be beneficial to embedded system design. First, embedded systems could make use of hybrid cooling with a Peltier thermo-electric cooler for higher performance and passive heat pipe conduction and radiation (see Resources). Similarly, an embedded design that is power-consumption sensitive most certainly should take advantage of DVS to reduce the CPU clock rate and Vsupply and thus reduce power consumption, much as the z990 does. Using DVS cycle rate reduction as a failsafe for reliability when device temperatures climb too high is a concept that could be beneficial to embedded systems. So, while achieving the peak performance of the z990 might be less important for embedded SoCs, the fail-safe thermal DVS design is definitely worth close study.
Figure 1. Z990 refrigeration cycle
Figure 1 shows the refrigeration cycle for the z990 MRUs (Modular Refrigeration Units). The cycle is adjusted based upon a closed loop with thermistor readings to operate in a high or low heat load mode. Refrigeration cycles like the z990's obey the first law of thermodynamics for a control volume. The enthalpy (internal energy storage per unit mass + Pressure x specific volume) is an important characteristic of the working fluid or refrigerant. Refrigeration cycles like the z990's move heat from an exchanger (evaporator) into the working fluid and then to the ambient atmosphere. The working fluid is cooled through compression and expansion; thus, work input is required to run the compressor (note that compressor must work harder when the ambient temperature is higher). The cycle allows the z990 MCMs to be cooled to match room temperature; passive cooling methods, in contrast, can only achieve heat flow with a significant temperature differential between the device and the ambient environment.
Heat pipe is a new cooling technology that has been used for exotic applications such as infrared instrument cooling on space telescopes, nuclear reactor cooling, and numerous sophisticated systems requiring passive heat transfer that is both highly reliable and highly efficient. Space-based applications have used heat pipe to transfer heat from solar cells exposed to the sun to radiators on the cool side of satellites facing deep, dark space.
Heat pipe almost appears to provide active cooling with no work input -- a violation of the first law of thermodynamics. Upon closer inspection, heat pipe is a scheme for very high thermal conductivity that uses a working fluid and the phenomena of fluid wicking to transport heat. Unlike convection, heat pipe requires no power input, and it achieves thermal conductivity that is unrivaled. The main barrier to the adoption of heat pipe has been cost. With improved production methods, heat pipe found its way into high- performance laptops when those laptops started generating so much heat that they actually posed a safety threat to users. IBM® used heat pipes to conduct heat away from CPUs (and user laps!) to keyboard and display hinge radiators in the ThinkPad model A20, introduced in 1999, making the A20 one of the coolest and most quiet laptops on the market at this time.
The growing use of heat pipe in commercial applications has helped to bring cost down and makes this once exotic heat transfer device a real possibility for many embedded systems. For more on this technology, see the links in the Resources section below.
Just as the z990 uses cycle steering (clock rate control and Vsupply), embedded systems such as the PowerPC 405LP use DVS to automatically adjust the Pswitching component of CMOS circuit power consumption. Since this term of the power consumption equation (see Equation 1) scales quadratically with Vsupply, significant power savings can be realized with only a slight decrease in clock rate and voltage. An application can detect that the workload has decreased based upon software performance monitoring and then lower the clock rate during idle or lower workload periods to conserve power. The simplest DVS interfaces provide a register interface and table of clock rates that software can select to conserve power. The DVS circuitry adjusts Vsupply for the lower clock rate selected by software in order to realize the power savings.
Laptop users are familiar with power management schemes out of necessity. Power management software must -- when possible -- spin down hard drives, turn off LCDs, and even hibernate and power down the CPU if a laptop is to have any reasonable battery lifetime. For the most part, power management is concerned with turning off currently unused peripherals and hibernating the CPU in a low power state as a last resort.
Scalable architectures such as PCI Express provide for peripheral power management for desktop and server systems as well. Here, the concern is good environmental policy and preventing the waste of power on idle computing devices. Embedded systems typically have few high-power peripherals, but nevertheless could still benefit from power management schemes such as those found in PCI Express (see Resources).
Users control laptop management of power, and simple formulas can configure machines to both maximize battery life and take advantage of grid power when available -- but embedded systems power management might require much more intelligence. For example, real-time systems, often embedded, must guarantee response to an external stimulus within a deadline. Research to integrate real-time embedded requirements with power management has just begun; for more on these issues, read the article "Maximizing the system value while satisfying time and energy constraints" (see Resources).
|Priority Example||Power Usage||Cooling||Performance||MIPS/watt|
|PowerPC G4 Laptop||Medium||High||Medium||Medium|
The z990 is distinguished by unparalleled performance and high availability and reliability. The zSeries® has been designed to operate with zero downtime, the ultimate in high availability. Cooling is important for high-availability systems because systems that run hot often have reliability issues as well as lower CPU performance. So, for the z990, the trade-off was made to incorporate active refrigeration and convective cooling to maximize performance and reliability. The cost and added complexity of active refrigeration might be prohibitive for embedded applications, but for a system like the z990 with zero downtime and high-performance design goals, active cooling is invaluable.
By comparison, a laptop, such as a notebook computer running the PowerPC G4, must better balance power, cooling, and performance. Cooling is a safety and usability issue for laptops. While active refrigeration and TEC cooling are impractical for laptops, simple convective fans combined with sophisticated heat pipe conduction and keyboard and display hinge radiation are fundamental. Laptops don't need performance that rivals the z990, of course, but must be competitive and must definitely exceed simpler mobile devices such as PDAs.
Finally, an embedded SoC such as the IBM 405LP is designed for long-life mobile and deeply embedded applications where power usage must be minimized. When power usage is minimized, cooling is not a significant issue, since much less heat will be generated in the first place. The idea of SoCs with DVS is to provide performance when needed by applications and to save power to the greatest extent possible at other times. Rather than providing raw MIPS like the z990 or even a G4 laptop, the 405LP is intended to provide the best performance per watt consumed.
System architects should carefully consider the relative importance of power, cooling, and performance for their specific application, as well as cost, reliability, and usability issues. Emergent power-conserving technologies and heat transfer methods will benefit all systems, ultimately leading to cooler systems with higher throughput and lower power consumption. But a perfect balance among the three factors is not always advisable, and the z990 and 405LP provide good examples of systems favoring performance or power consumption, respectively.
- The hybrid cooling and cycle steering design of the z990
described in "Hybrid cooling with cycle steering in the IBM eServer
z990," G. F. Goth, D. J. Kearney, U. Meyer, and D. W. Porter
(IBM Journal of Research and Development, 2004) is a
sophisticated active cooling approach that optimizes performance with
low power consumption. The design is an enhancement of earlier
active refrigeration designs developed by IBM for the zSeries of
- The first IBM S/390 G4 system to use active refrigeration for
higher performance and lower power consumption is described in "High-
end server low-temperature cooling," R. R. Schmidt and B. D.
Notohardjono (IBM Journal of Research and Development, 2002, PDF).
- The Thermacore company provides a nice overview on the physics of heat pipes. Heat pipes
were first invented at Los Alamos National Laboratories and have
found use in many deep space and satellite systems for conducting
heat efficiently between hot and cold surfaces, as well as spin-off
applications in commercial computing such as laptop computer cooling.
IBM has developed novel methods using heat pipes to conduct heat
from processors to radiating surfaces as described "Cooling off laptops," Tom Staudter
(Think Research, August 2001).
47 of the IBM Journal of Research and Development is
dedicated to power-efficient computing technology and describes power
efficiency research projects at IBM Austin.
- The article, "Maximizing the system value while satisfying time and
energy constraints," C. A. Rusu, R. Melhem, and D. Mosse (IBM
Journal of Research and Development, 2003) provides a framework
and scheme for software-managed trade-offs between power and
performance using voltage scaling.
- "Frequency switching improves power management in Power
Architecture chips," Helena Purgatorio (developerWorks,
September 2004) provides a great review of PowerPC power management
- Thermo-electric Peltier coolers can provide localized cooling
with the cost of some additional power to drive the cooling element.
In cases where cooling is more important than total power
consumption, a TEC might be very appropriate. Learn more in "Keeping cool," Robert Otey and Barry Moskowitz
(OE Magazine, March 2001).
- Find numerous additional articles on various cooling methods, design,
and technologies in Electronics Cooling.
- Design tools such as the PowerTimer toolset can be used to design
configurable processor core ASICs and SoCs for power sensitive
applications. Learn more in "New methodology for early-stage,
microarchitecture-level power-performance analysis of
microprocessors," D. Brooks, P. Bose, V. Srinivasan, M. K.
Gschwind, P. G. Emma, and M. G. Rosenfield (IBM Journal of
Research and Development, 2003).
- Many SoC embedded processors, including the IBM PowerPC 405LP,
now have dynamic voltage and frequency scaling. "Dynamic
Power Management for Embedded Systems IBM and MontaVista Software,"
a white paper from IBM and MontaVista Software, discusses the OS
interface to power management and cycle steering features.
- The Peripheral Component Interconnection bus, PCI, includes a
power management scheme for scalable systems described in the PCI Express
specification. See also Peter Seebach's Standard and specs
column on "The PCI bus" (developerWorks, February 2005).
- Get a nice overview of the laws of thermodynamics
that affect cooling system designs, including the most
fundamental zeroth law.
- For a look at the lighter side of cooling issues, see the Power Architecture challenge, "Cooling down hot processors" (developerWorks, February 2005).
- The Power Architecture Community Newsletter includes full-length articles as well as recent news about members of the Power Architecture community and upcoming events of interest.
about the Power Architecture Community Newsletter and how to contribute to it. Subscription is free.
Dr. Sam Siewert is an embedded system design and firmware engineer who has worked in the aerospace, telecommunications, and storage industries. He also teaches at the University of Colorado at Boulder part-time in the Embedded Systems Certification Program, which he co-founded. His research interests include autonomic computing, firmware/hardware co-design, microprocessor/SoC architecture, and embedded real-time systems.