CMOS power dissipation VL9252
Content
CHAPTER 3
CMOS POWER DISSIPATION
3.1 INTRODUCTION
The CMOS power dissipation has become a very hot topic during the last decade or so. The number of battery-powered hand-held applications, e.g. mobile phones and laptop computers is steadily increasing and more and more functions are integrated into the systems, e.g. multi-media applications in mobile phones. This is one of the driving forces for analysis of the mechanisms of power dissipation and power-reduction techniques. Another driving force is the incredible power dissipation of state-of-the-art microprocessors where heat removal and current delivery are very hard and expensive to accomplish.
3.2 MECHANISMS OF POWER DISSIPATION
Mechanisms of power dissipation are usually divided into two classes: dynamic and static power dissipation. Dynamic power dissipation occurs when the circuit is operational, i.e. the circuit is performing some task on some data. Static power dissipation becomes an issue when the circuit is inactive or in a power-down mode.
3.2.1 Dynamic Power Dissipation
Dynamic power dissipation can be further subdivided into three mechanisms: switched, short-circuit, and glitch power dissipation. All of them more or less depend on the activity, timing, output capacitance, and supply voltage of the circuit. The repeated charging and discharging of the output capacitance is necessary to transmit information in CMOS circuits. This charging and discharging causes for the switched power dissipation. The power consumption of a CMOS digital circuit can be represented as P = fCVdd2 + f I short Vdd + I leak Vdd
(3.1)
Where f is the clock frequency , C is the average switched Capacitance per clock cycle, Vdd is the supply voltage , I
short
is the short circuit current and I
leak
is the
leakage current. In a well optimized low power VLSI circuits, the Ist term of equation (3.1) is by far the dominant. The stand by power consumption is accounted for by the 3rd term. Using a lower Vdd is an effective way to reduce the dynamic power consumption since Ist term is proportional to the square of Vdd.It should also be noted that the short circuit and leakage power dissipation are also strongly dependent on Vdd.However, using a lower Vdd degrades performance.
3.2.2 Short-Circuit Power Dissipation
In real circuits signals have non-zero rise and fall times which causes both the P net and the N net of the CMOS gate to conduct current simultaneously.
This leads to the flow of a short-circuit current for a short period of time. The input and output slopes of a gate should be equal to minimize the overall shortcircuit dissipation in gates (John Rabaey (2003)).Also large load capacitance can significantly reduce the short-circuit dissipation of the driving gate (John Rabaey (2003)),(Weste.N and Eshraghian.K (1993) ).
3.2.3 Glitch Power Dissipation
Glitches are undesired signal transitions which do not contribute any useful information. Still they cause switched and short-circuit power dissipation. Glitches can be divided into two categories: generated and propagated. If the input signals to a gate are skewed in time, there is a clear and present danger of having a generated glitch at the output. If a glitch arrives at the input of a gate and if the input is sensitive at the moment, a propagated glitch will be created .The number of glitches in a circuit, comprising many gates, depends on the logic depth, the logic function and the gate fanouts. In some circuits, the major part of the dynamic power dissipation is due to glitches. Since glitches sometimes have peak voltages that are in the middle of the transition interval, their contribution to the shortcircuit power dissipation can be non-negligible (John Rabaey (2003)).However, the glitch can be minimized by scale down the supply and threshold voltages probably at different phases. The rates at which these voltages are scaled depend on which type of process it is low standby, low power, or high performance.(Ayman A et al (2001)) The glitch behavior under these
circumstances has been studied where two voltage-scaling techniques have been used. One technique where the threshold voltage is kept constant when the supply voltage is lowered and the second technique where the threshold voltage is scaled at the same rate as the supply voltage.
3.2.4 Static Power Dissipation
The static power components become important when the circuits are at rest, i.e. when there is no activity in the circuits and they are all biased to a specific state. The static power dissipation includes sub threshold and reversedbiased diode leakage currents. Due to the necessary but harmful (in a leakagepower sense) down-scaling of threshold voltages, the sub threshold leakage is becoming more and more pronounced. Below the threshold voltage, in weak inversion, the transistors are not completely off. The sub threshold current has a strong dependence on the threshold voltage.
3.3 Reduction of Power Dissipation
In the beginning of the last decade, battery-powered hand-held devices such as mobile phones and laptop computers emerged. This called for low-power operation and a lot of design methods on different hierarchy levels were summarized by Chandrakasan et al. (Chandrakasan.C (1992)).
3.3.1 Voltage Scaling and Reduced Voltage Swing
Reducing the supply voltage is an attractive solution to reduce the power dissipation since both the switched and the short-circuit power dissipation
have a strong Vdd dependence. A delay penalty can be mitigated by reducing the threshold voltage but then the sub threshold leakage will increase exponentially. Hence it is important to select appropriate threshold and supply voltages. Another way of sustaining the throughput is to do an architectural voltage scaling. Then the throughput is kept by either parallelization (Ayman A et al (2001)) or pipelining (Pucknell (2004)).Both techniques introduce some overhead switched capacitance due to the extra hardware added but this is power-wise compensated for by the lower supply voltage. For long buses with high capacitance, a reduced voltage swing can be a solution.
To reduce the active (dynamic) power, Vdd and Vth (Tezaswi Raja et al (2004)),(Uming Ko et al (1995)) hopping schemes have been proposed. In hopping scheme there is cooperation between the hardware and the software, and the software controls the supply voltage of the circuit and adapts the voltage to the present task. Two or more voltages can be used but the drawback is that it takes some time for the circuit to adjust to the new voltage. The threshold voltages are controlled by back-biasing. In modern processes, the uses of at least two threshold voltages are common. One low Vth for the delay-critical part of the circuit and a higher one for the rest. Since both delay and dynamic and sub threshold dissipations depend on both sizing and threshold voltages, there is, for delayconstrained circuits, a trade-off between using a smaller transistor with lower Vth and larger transistor with higher Vth and this trade-off depends on the activity of the circuit.
3.3.2 Clock Frequency Reduction
Reducing the clock frequency is not as beneficial as reducing the supply voltage. However, many processors of today have different power-down modes where the clock signal is silenced to blocks of the application that are not used at the moment. This is referred to as clock gating. Clock gating can in most cases be used in conjunction with other low-power techniques.
3.3.3 Switched Capacitance Reduction
Reducing the switched capacitance is as efficient as reducing the clock frequency. Many techniques have been proposed to reduce the switched capacitance. The selection of logic style can significantly affect the critical capacitances. When a low-power solution is searched for, conventional static CMOS is often a safe bet. Multiplexers and XOR gates are an exception since they can be implemented in pass-transistor logic styles using fewer transistors. In flipflops and registers, the capacitances of the clock nodes are important since the clock signal has a high activity. Therefore, flip-flops with a small number of clocked transistors have been proposed (Weste.N and Eshraghian.K (1993) ).
CMOS POWER DISSIPATION
3.1 INTRODUCTION
The CMOS power dissipation has become a very hot topic during the last decade or so. The number of battery-powered hand-held applications, e.g. mobile phones and laptop computers is steadily increasing and more and more functions are integrated into the systems, e.g. multi-media applications in mobile phones. This is one of the driving forces for analysis of the mechanisms of power dissipation and power-reduction techniques. Another driving force is the incredible power dissipation of state-of-the-art microprocessors where heat removal and current delivery are very hard and expensive to accomplish.
3.2 MECHANISMS OF POWER DISSIPATION
Mechanisms of power dissipation are usually divided into two classes: dynamic and static power dissipation. Dynamic power dissipation occurs when the circuit is operational, i.e. the circuit is performing some task on some data. Static power dissipation becomes an issue when the circuit is inactive or in a power-down mode.
3.2.1 Dynamic Power Dissipation
Dynamic power dissipation can be further subdivided into three mechanisms: switched, short-circuit, and glitch power dissipation. All of them more or less depend on the activity, timing, output capacitance, and supply voltage of the circuit. The repeated charging and discharging of the output capacitance is necessary to transmit information in CMOS circuits. This charging and discharging causes for the switched power dissipation. The power consumption of a CMOS digital circuit can be represented as P = fCVdd2 + f I short Vdd + I leak Vdd
(3.1)
Where f is the clock frequency , C is the average switched Capacitance per clock cycle, Vdd is the supply voltage , I
short
is the short circuit current and I
leak
is the
leakage current. In a well optimized low power VLSI circuits, the Ist term of equation (3.1) is by far the dominant. The stand by power consumption is accounted for by the 3rd term. Using a lower Vdd is an effective way to reduce the dynamic power consumption since Ist term is proportional to the square of Vdd.It should also be noted that the short circuit and leakage power dissipation are also strongly dependent on Vdd.However, using a lower Vdd degrades performance.
3.2.2 Short-Circuit Power Dissipation
In real circuits signals have non-zero rise and fall times which causes both the P net and the N net of the CMOS gate to conduct current simultaneously.
This leads to the flow of a short-circuit current for a short period of time. The input and output slopes of a gate should be equal to minimize the overall shortcircuit dissipation in gates (John Rabaey (2003)).Also large load capacitance can significantly reduce the short-circuit dissipation of the driving gate (John Rabaey (2003)),(Weste.N and Eshraghian.K (1993) ).
3.2.3 Glitch Power Dissipation
Glitches are undesired signal transitions which do not contribute any useful information. Still they cause switched and short-circuit power dissipation. Glitches can be divided into two categories: generated and propagated. If the input signals to a gate are skewed in time, there is a clear and present danger of having a generated glitch at the output. If a glitch arrives at the input of a gate and if the input is sensitive at the moment, a propagated glitch will be created .The number of glitches in a circuit, comprising many gates, depends on the logic depth, the logic function and the gate fanouts. In some circuits, the major part of the dynamic power dissipation is due to glitches. Since glitches sometimes have peak voltages that are in the middle of the transition interval, their contribution to the shortcircuit power dissipation can be non-negligible (John Rabaey (2003)).However, the glitch can be minimized by scale down the supply and threshold voltages probably at different phases. The rates at which these voltages are scaled depend on which type of process it is low standby, low power, or high performance.(Ayman A et al (2001)) The glitch behavior under these
circumstances has been studied where two voltage-scaling techniques have been used. One technique where the threshold voltage is kept constant when the supply voltage is lowered and the second technique where the threshold voltage is scaled at the same rate as the supply voltage.
3.2.4 Static Power Dissipation
The static power components become important when the circuits are at rest, i.e. when there is no activity in the circuits and they are all biased to a specific state. The static power dissipation includes sub threshold and reversedbiased diode leakage currents. Due to the necessary but harmful (in a leakagepower sense) down-scaling of threshold voltages, the sub threshold leakage is becoming more and more pronounced. Below the threshold voltage, in weak inversion, the transistors are not completely off. The sub threshold current has a strong dependence on the threshold voltage.
3.3 Reduction of Power Dissipation
In the beginning of the last decade, battery-powered hand-held devices such as mobile phones and laptop computers emerged. This called for low-power operation and a lot of design methods on different hierarchy levels were summarized by Chandrakasan et al. (Chandrakasan.C (1992)).
3.3.1 Voltage Scaling and Reduced Voltage Swing
Reducing the supply voltage is an attractive solution to reduce the power dissipation since both the switched and the short-circuit power dissipation
have a strong Vdd dependence. A delay penalty can be mitigated by reducing the threshold voltage but then the sub threshold leakage will increase exponentially. Hence it is important to select appropriate threshold and supply voltages. Another way of sustaining the throughput is to do an architectural voltage scaling. Then the throughput is kept by either parallelization (Ayman A et al (2001)) or pipelining (Pucknell (2004)).Both techniques introduce some overhead switched capacitance due to the extra hardware added but this is power-wise compensated for by the lower supply voltage. For long buses with high capacitance, a reduced voltage swing can be a solution.
To reduce the active (dynamic) power, Vdd and Vth (Tezaswi Raja et al (2004)),(Uming Ko et al (1995)) hopping schemes have been proposed. In hopping scheme there is cooperation between the hardware and the software, and the software controls the supply voltage of the circuit and adapts the voltage to the present task. Two or more voltages can be used but the drawback is that it takes some time for the circuit to adjust to the new voltage. The threshold voltages are controlled by back-biasing. In modern processes, the uses of at least two threshold voltages are common. One low Vth for the delay-critical part of the circuit and a higher one for the rest. Since both delay and dynamic and sub threshold dissipations depend on both sizing and threshold voltages, there is, for delayconstrained circuits, a trade-off between using a smaller transistor with lower Vth and larger transistor with higher Vth and this trade-off depends on the activity of the circuit.
3.3.2 Clock Frequency Reduction
Reducing the clock frequency is not as beneficial as reducing the supply voltage. However, many processors of today have different power-down modes where the clock signal is silenced to blocks of the application that are not used at the moment. This is referred to as clock gating. Clock gating can in most cases be used in conjunction with other low-power techniques.
3.3.3 Switched Capacitance Reduction
Reducing the switched capacitance is as efficient as reducing the clock frequency. Many techniques have been proposed to reduce the switched capacitance. The selection of logic style can significantly affect the critical capacitances. When a low-power solution is searched for, conventional static CMOS is often a safe bet. Multiplexers and XOR gates are an exception since they can be implemented in pass-transistor logic styles using fewer transistors. In flipflops and registers, the capacitances of the clock nodes are important since the clock signal has a high activity. Therefore, flip-flops with a small number of clocked transistors have been proposed (Weste.N and Eshraghian.K (1993) ).
