Power Supply Aware Computing 电源感知计算

Power Supply Aware Computing

电源感知计算

摘要：本文讨论了与电源通信的微处理器的性能效益和能效改进。电源对计算负载变化的响应可以通过关于预期瞬态的信息大大改善。可以修改电源电压，以减少稳态功率损失。微处理器还可以根据电源提供的信息来调整其计算活动。还有的好处，如降低成本、缩小尺寸和提高可靠性。实验结果证明了其优点

Abstract—This paper discusses the performance benefits and

energy efficiency improvements from microprocessors that

communicate with their power supply. The power supply's

response to computing load changes can be vastly improved with

information about expected transients. Supply voltages can be

modified to decrease steady-state power losses. A

microprocessor can also adjust its computational activity based

on information provided by the power supply. Additional

benefits such as cost reduction, size reduction, and improved

reliability are expected. Experimental results demonstrate the

Benefits

I. 介绍微处理器需要在低电压（约1 V)和大电流(100 A或更多）下的功率。由于处理负载步骤而导致的电流需求的变化速率可以是数百个A/µs，同时需要一个严格调节的电压供应（如50 mV最大偏差）[1]。如果电源电压偏离了给定的范围，微处理器就会出现处理错误。因此，降压转换器需要一个大的输出电容和一个低电感的电感器（如图1所示），以满足供电规格。

I.

INTRODUCTION

Microprocessors need power at low voltages (around 1 V)

and high currents (100 A or more at full load). The rate of

change in current demand due to processing load steps can be

hundreds of A/µs while simultaneously requiring a tightly

regulated voltage supply (such as 50mV maximum deviation)

[1]. If the supply voltage were to deviate outside the given

bounds, processing errors would occur in the microprocessor.

As a result, a large output capacitance and an inductor with

low inductance is necessary in the buck converter (shown in

Fig. 1) that typically supplies microprocessors in order to meet

the supply specifications.

为了克服电源的低输出阻抗要求，提高瞬态响应，多降压转换器通常并联连接。该多相降压转换器具有增加的有效控制带宽，可以降低输出电压纹波，其中为转换器中的降压相数。多相方法也分配了变频器的应力和散热，但扩大了部件数量，提高了成本。

To overcome the low output impedance requirement for

the power supply and to improve transient response, multiple

buck converters are commonly connected in parallel. This

multiphase buck converter has an increased effective control

bandwidth and can decrease output voltage ripple by a factor

of ݊ through switch interleaving, where ݊ is the number of

buck phases in the converter. A multiphase approach also

distributes converter stresses and heat dissipation but expands

parts count and raises cost.

转换器响应负载瞬变的能力往往受到所实现的控制方法的限制。采用线性控制方法，转换器的控制带宽一般为开关频率的十分之一（典型的电源开关在200-500 kHz）。这激发了非线性控制方法，如最小时间控制，可以应用于瞬态，以提高转换器性能[2]。

ability of the converter to respond to load transients is often limited by the control method implemented. With linear control methods, the control bandwidth of the converter is generally one-tenth of the switching frequency (typical supplies switch at 200-500 kHz). This has motivated nonlinear control methods such as minimum time control that can be applied during a transient to improve converter performance [2].

图1.微处理器的调压器通常采用降压转换器。可以增加增强分支数，以提高性能。

Figure 1. A buck converter is commonly used in voltage regulators for

microprocessors. Augmentation branches can be added to improve

performance.

如果微处理器和电源之间有更多的交互作用，电源设计人员面临的许多挑战可以被简化。在下面的章节中，我们将解释其中的一些好处。第二节讨论了在使用固定拓扑结构时，降压转换器性能的物理限制。在第三节中，重点是为电源提供信息的微处理器的优势。第四节研究了另一个范例，其中微处理器根据提供给它的关于电源状态的信息来调整其活动。

A number of challenges facing power supply designers

could be simplified if there was more interaction between the

microprocessor and the supply. In the following sections,

some of the benefits will be explained. Section II discusses the

physical limits to buck converter performance when a fixed

topology is used. In Section III, the focus is on advantages of

the microprocessor providing the power supply with

information. Section IV examines another paradigm where the

microprocessor adjusts its activity based on information given

to it about the state of the power supply.

微光电源控制逻辑电路的电源电压（Vdd）必须保持在一个固定的范围内。如果没有，逻辑元素之间的延迟时间会发生变化，导致锁存错误，从而可能引入错误计算。虽然有多种方法来减轻这些误差的影响，[3]，[4]，但通过严格调节的电源电压来避免这些误差是必要的。

II. POWER SUPPLY CONTROL

The supply voltage (V

dd) for logic circuits must be

maintained within a fixed range. If not, variations in the delay

time between logic elements occur resulting in latching errors

that can introduce miscalculations. While there are a variety of

methods to mitigate the effects of these errors [3], [4],

avoiding these errors with a tightly regulated supply voltage is

necessary.

调节标准线性控制的电源电压是一项相当简单的任务。当负载突然变化时，保证电压不偏离超出规定的范围是相当具有挑战性的。负载瞬态会导致电源电压的大过调或过调。为了解决这一问题，在硬件设计中，一个常见的做法是增加电源的降压转换器的输出电容。这提供了一个能量缓冲，可以利用。与这种方法相关的一些缺点是增加了成本和板脚率

Regulating the supply voltage in steady state is a fairly

simple task with standard linear controls. Guaranteeing that

the voltage does not deviate outside the specified bounds

when the load abruptly changes is considerably more

challenging. Load transients can induce large overshoot or

undershoot in the supply voltage. To combat this problem, a

common practice in hardware design is to increase the output

capacitance of the power supply’s buck converter. This

provides an energy buffer that can be utilized when load steps

occur. Some of the drawbacks associated with this approach

are increased cost and board footp

图2.降压转换器对负载上升的最小时间响应，无延迟或开关定时不准确。电压偏差是固定拓扑可能的最小量

Figure 2. Minimum time response of buck converter to a load step-up

with no delay or switch timing inaccuracies. Voltage deviation is the

minimum amount possible for a fixed topology

许多控制技术来改善电源的瞬态响应。作为一个电压源，其目标是在一个电流水平的范围内提供恒定的电压。降压转换器中的平均电感电流等于稳态负载电流，因此在瞬变期间，电感电流必须从一个电流水平移动到下一个电流水平。因此，降压转换器性能的一个基本限制是电感器的旋转率。旋转率与电感上的电压成正比，与其电感成反比。许多控制方法将在稳态运行中应用线性控制器，在瞬态运行中应用非线性控制器将转换器状态从一个工作点移动到另一个工作点。为了最小化电压偏差，需要一个快速的瞬态响应。

Numerous control techniques have been developed to

improve the transient response of the power supply. As a

voltage source, the goal is to supply constant voltage over a

range of current levels. The average inductor current in a buck

converter is equal to the steady-state load current, so the

inductor current must move from one current level to the next

during transients. Hence, one of the fundamental limits to the

performance of a buck converter is the inductor slew rate. The

slew rate is proportional to the voltage across the inductor and

inversely proportional to its inductance. Many control

methods will apply a linear controller during steady-state

operation and a nonlinear controller to move the converter

states from one operating point to the next during a transient.

A swift transient response is necessary in order to minimize

voltage deviation.

最小时间控制，也称为时间最优控制，可用于将具有固定拓扑结构的功率转换器的瞬态响应推到其物理极限[5]-[7]。结果表明，用一组开关动作[8]-[10]可以在最短的时间内达到一个新的操作点。对于图2中所示的负载上升，高侧开关（q1）将首先打开，通过电感器的电流将增加。输出电容器提供负载电流和电感电流之间的差，因此电容器电压首先下降。电感器电流必须超过新的负载电流水平，以取代电容器中丢失的电荷。在适当的时间通过控制关闭高侧开关，使损失和替换的电荷（分别用区域Q1和Q2表示）相等。

Minimum time control, also known as time-optimal

control, can be used to push the transient response of a power

converter with a fixed topology to its physical limits [5]-[7]. It

has been shown that a new operating point can be reached in

minimum time with one set of switch actions [8]-[10]. For the

load step-up depicted in Fig. 2, the high side switch (q1) will

first turn on and the current through the inductor will increase.

The output capacitor supplies the difference between the load

current and inductor current, so the capacitor voltage dips at

first. The inductor current must go beyond the new load

current level to replace the charge in the capacitor that was

lost. The high side switch is turned off by control at the

appropriate time such that the charges lost and replaced

(represented by area Q1 and Q2, respectively) are equal.

对于实现最小时间控制有几个挑战。其中一个关键因素是准确和快速地检测负载步骤。当电容器电压超过一个电压阈值时进行传感是一种常见的方法，但这将在响应中引入一个延迟。在检测到瞬态后，必须根据负载步长精确地确定开关动作的时间，以确保时间最优性。这可能需要在电压调节器的控制器中进行大量的处理，因此限制了开关频率。与最小时间控制相关的理想波形如图2所示。在这方面也有一些困难

There are several challenges with implementing minimum

time control. One key element is detecting load steps

accurately and swiftly. Sensing when the capacitor voltage

passes a voltage threshold is a common approach, but this

introduces a delay in the response. After the transient is

detected, the timing of the switching action must be precisely

determined based on the load step size to ensure time

optimality. This may require significant processing in the

voltage regulator’s controller and therefore limits switching

frequency. Ideal waveforms associated with minimum time

control are shown in Fig. 2. There is also some difficulty in

从瞬态控制方法平稳地过渡到稳态控制器。罗马数字 3通信电源和微处理器通常作为单独的系统进行设计和操作。电源设计师得到了微处理器制造商[1]的要求和期望清单。在运行期间，两个系统之间确实发生的有限的交互作用主要包括状态指示器和稳态参考信号（例如，电压识别数字，或VID）

smoothly transitioning from the transient control method to the steady-state controller. III. OMMUNICATION Power supplies and microprocessors are generally designed and operated as separate systems. Power supply designers are given a list of requirements and expectations by microprocessor manufacturers [1]. The limited interaction that does occur between the two systems during operation consists primarily of status indicators and steady-state reference signals (e.g. Voltage Identification Digital, or VID)

通过微处理器和电源之间的更多交互，微处理器电源面临的许多挑战可以减轻。不仅可以提高性能，还可以提高能源效率、可靠性和成本。该微处理器也可以受益于来自电源的额外信息。A. 微处理器与电源通信

Many of the challenges facing microprocessor power  supplies could be mitigated with more interaction between the microprocessor and the power supply. Not only could there be improvements to performance but also energy efficiency, reliability, and cost. The microprocessor may also benefit from additional information from the power supply. A. Microprocessor communicating with the power supply

如图3所示，以提高性能和效率。例如，如果微处理器能够提供关于负载步长、方向或定时的预先通知，则电源可以更快速和准确地响应。事实上，甚至包括这些信息中的一条，都有可能大大改善瞬态响应。将消除非线性控制器中的传感和计算延迟，并可以采取行动。这将导致电源电压中的过调或过调减少。电压过调特别难以管理，因为在低侧开关打开的降压转换器中，穿过电感器的电压近似等于输出电压。当高侧开关被激活时，电感器上的电压是输入电压和输出电压之间的差值。这意味着电感器在负载减少时的旋转率可以比负载增加时慢一个数量级

wer supply, as suggested in Fig. 3, to improve performance and efficiency. For example, if the microprocessor could provide advance notice of the load step size, direction, or timing, the power supply could respond more swiftly and accurately. In fact, including even one of these pieces of information has the potential to substantially improve the transient response. The sensing and computation delay in a nonlinear controller would be removed and action could be taken. This would lead to reduced overshoot or undershoot in the supply voltage. Voltage overshoot is especially difficult to manage since the voltage across the inductor in a buck converter with the low side switch on is approximately equal to the output voltage. When the high side switch is activated, the voltage across the inductor is the difference between the input voltage and the output voltage. This means that the slew rate of the inductor during a load decrease can be an order of magnitude slower than during a load increase

虽然对于一个具有最小时间控制的固定拓扑降压转换器，已经达到了瞬态响应的物理极限，但甚至有可能有更好的性能。转换器增强算法可以通过增加额外的能量路径[11]-[13]来消除负载瞬态。增强路径将在负载增加时提供能量，或在负载减少时通过贡献负载电流和电感器电流之间的差值来吸收能量。虽然用于增强的组件可以很简单，

While the physical limits of transient response are reached for a fixed topology buck converter with minimum time

control, it is possible to have even better performance. Converter augmentation has been proposed to nullify load transients through adding extra energy paths [11]-[13]. The augmentation paths will either supply energy during a load increase or sink energy during a load decrease by contributing the difference between the load current and inductor current. Although the components used in augmentation can be simple,

图3：微处理器可以通知电源。Figure 3. The microprocessor can inform the power supply when a load step will occur and how large the transient will be.

由于传感延迟，它可能很难实现。如果微处理器提供了负载步骤信息，转换器的增强将被简化。此外，如果提供预先信息，可以在预期载荷步骤时改变电感能量。这将支持负载前馈效应，使电源能够主动限制由于负载瞬态引起的电压偏差。虽然之前已经提出了一种前馈方法，即[14]，但它依赖于额外的传感电路来重建负载电流估计。

it can be difficult to implement because of sensing delays. If the microprocessor provided the load step information, converter augmentation would be simplified. Furthermore, the inductor energy could be altered in anticipation of a load step if advance information is provided. This would support a load feedforward effect that enables the power supply to proactively limit the voltage deviation due to load transients

While a feedforward approach has been suggested previously [14], it has relied upon additional sensing circuitry to reconstruct load current estimates.

电源通常使用一种称为负载线调节或下垂（通常称为自适应电压定位）的技术，以允许更多的电压摆动净空间来响应负载步骤。在轻负载时，电压保持在可接受电压带的顶部附近，而在重负载时，电压保持在电压带的底部附近。这就可以充分利用电压带。或者，如果瞬态响应改善到不再需要这个净空空间，那么稳态电压可以保持在电压带的底部附近。这可以适应在轻负载下降低电压约100 mV，基于平均50%的负载，可以将微处理器的功耗降低约10%

Power supplies normally use a technique called load-line regulation or droop (often termed adaptive voltage

positioning) to allow more voltage swing headroom for responding to load steps. At light loads the voltage is

maintained near the top of the acceptable voltage band, and it is maintained near the bottom of the band at heavy loads. This enables full utilization of the voltage band. Alternatively, if the transient rsponse was improved to the point that this headroom were no longer necessary, the steady-state voltage could remain near the bottom of the voltage band. This might accommodate lowering the voltage by about 100 mV at light loads which could reduce microprocessor power consumption by about 10%, based on average 50% loading

B. 与微处理器通信的电源通常，电源必须调节到给定的参考（减少下垂），并支持微处理器负载。在大多数情况下，微处理器认为供电是理所当然的，并且在运行时几乎不考虑供电条件。这种模式可以逆转，这样微处理器将根据电源要求的限制限制其活动（如图4所示）。这主要适用于微处理器想要改变其计算负载时。该微处理器可以根据电源的旋转率来调整其计算负载的斜坡率。目标是匹配旋转速率，使电源电压（Vdd）没有偏差。

B. Power supply communicating with the microprocessor

Generally the power supply must regulate to a given reference (less droop) and support the microprocessor load. For the most part, the microprocessor takes the supply for granted and operates with little consideration of power supply conditions. This paradigm could be reversed such that the microprocessor would throttle its activity based on limits requested by the power supply (illustrated in Fig. 4). This applies primarily when the microprocessor wants to change its computational load. The microprocessor could adjust the ramp

rate of its computational load based on the slew rate of the power supply. The objective would be to match slew rates so that there would be no deviation in the supply voltage (Vdd).

设施将在短时间内受到限制，但会改善电源的效率和电压调节。由于微处理器的旋转率比电源大一个数量级，所以它可以很容易地匹配电源的旋转率。这种方法的一个挑战将是在负载减少期间。微处理器可能需要知道一组计算何时有可能提前完成，这样它就可以逐渐减少其计算活动，而不是突然结束。微处理器将发出电源供应的信号

ility would be limited for short intervals, but there would improvements in the power supply's efficiency and voltage regulation. Since the slew rate of the microprocessor is orders of magnitude greater than the power supply, it can easily match the power supply’s slew rate. One challenge with this approach would be during a load decrease. The microprocessor would likely need to know when a set of calculations is likely to complete ahead of time so that it can gradually decrease its computational activity instead of ending

abruptly. The microprocessor would signal the power supply 负载正在减少，然后跟踪电源的输出电流的减少速率。考虑到零部件数量和成本的潜在减少，很可能会有系统效益。由于电源电压超冲和欠冲可以用这种方法调节，因此需要更少的散装电容器。这将导致提高可靠性。

that the load is decreasing and then track the power supply’s rate of decrease in output current. There are likely to be system benefits, given the potential reduction in parts count and cost. Since supply voltage overshoots and undershoots can be moderated with this approach, fewer bulk capacitors would be required. This would lead to improvements in reliability.

增值为了证明电源感知计算的好处，我们给出了关于负载步长、方向和时间的先验信息的增强降压转换器。这模拟了与电源通信的微处理器将表现出的一种行为。负载前馈电路使降压转换器可以直接从一个操作点移动到下一个操作点，而在电源电压中几乎没有超调或下调。与商用多相转换器相比，这种特殊的转换器是一个具有相对较高阻抗和最小输出电容的测试单元。对它的测试代表了用这种控制方法可以实现的相对改进。

IV. EXPERIMENTAL EXAMPLES

To demonstrate the benefits of power supply aware  computing, an augmented buck converter given prior

information about load step size, direction, and timing is examined. This mimics one type of behavior that a

microprocessor that communicates with the power supply would exhibit. The load feedforward enables the buck converter to move directly from one operating point to the next with little overshoot or undershoot in the supply voltage. This particular converter is a test unit with relatively high impedance and minimal output capacitance compared to commercial multi-phase converters. Tests on it represent relative improvements that can be achieved with this control approach.

图5 (a)中对60%负载下降的瞬态响应显示了小于50 mV的电压超调。电感器电流直接移动到增强降压转换器中的新负载电流值。尽管从一个操作点到下一个操作点的过渡时间是缓慢的20-30µs，

The transient response to a sixty percent load step-down in Fig. 5 (a) displays a voltage overshoot of less than 50 mV. The inductor current moves directly to the new load current value in the augmented buck converter. Even though the transition time from one operating point to the next is a slow 20-30 µs,

具有已知的负载步长和时间。(b)负载逐步升降在一个增强降压转换器与已知的负载步长和时间

with known load step size and time. (b) Load steps up and down in an augmented buck converter with known load step size and tim

他对电压调节的干扰是最小的。增强分支在瞬态运行期间去除多余的能量，并在稳态运行时保持关闭。如果在同一转换器中实现最小时间控制，电感电流将需要降低到新的负载水平之前，然后回到新的负载水平。这使过渡时间增加了一倍，电压超调时间大约增加了四倍。在图5 (b)中，时间尺度被扩展，并观察到电流对几个负载步长的急剧响应。随着增强和负载步长信息的增加，电源电压的瞬态偏差几乎消失。

he disruption in voltage regulation is minimal. The augmentation branch removes the excess energy during the transient and remains off during steady-state operation. If minimum time control were implemented in this same converter, the inductor current would need to decrease beyond the new load level before coming back to the new operating point. This doubles the transition time and roughly quadruples the voltage overshoot. In Fig. 5 (b) the time scale is expanded and a sharp current response to several load steps in current is observed. The supply voltage transient deviations are nearly nullified with augmentation and load step information.

V.结论微处理器的电源要求在电压调节和负载电流旋转率方面是严格的。为了避免可能导致处理错误的电源电压偏差，在电源中通常采用大体积电容。这增加了成本，增加了主板占用的空间，并降低了可靠性。虽然已经发展了各种非线性控制技术，可以快速地从一个载荷条件移动到下一个载荷条件，但也遇到了瞬态响应的物理旋转极限

V. CONCLUSION

Power supply requirements for microprocessors are stringent in terms of voltage regulation and load current slew rates. To avoid supply voltage deviation which could lead to processing errors, a large bulk capacitance is commonly used in power supplies. This increases cost, increases motherboard footprint, and reduces reliability. Although a variety of nonlinear control techniques have been developed to swiftly move from one load condition to the next, the physical slewrate limits of transient response are encountered本文中提出的另一种范例包括增加微处理器和电源之间的通信。如果给电源关于负载步长和定时的信息，则电源电压瞬态几乎无效。另一种方法是根据电源的旋转速率来限制微处理器的活动。这将减轻电源电压的偏差，同时只是暂时限制计算活动的变化。实验结果表明，一个瞬态响应几乎无效的增强功率转换器。

An alternative paradigm presented in this paper involves

increased communication between the microprocessor and

power supply. Supply voltage transients can be nearly

nullified through augmentation if the power supply is given

information about the load step size and timing. Another

approach is to throttle the activity of the microprocessor based

on the slew rate of the power supply. This would mitigate

supply voltage deviation while only temporarily constraining

the change in computational activity. Experimental results that

demonstrate a nearly nullified transient response have been

presented for an augmented power converter.

作者要感谢M. Sweeney和V. Bora在设计、建造和测试增强降压转换器方面的帮助。参考[1]英特尔公司，“电压调压器模块（VRM）和企业电压调压器下降（EVRD）11.1设计指南”，

ACKNOWLEDGMENTS

The authors would like to thank M. Sweeney and V. Bora

for their assistance with designing, building, and testing the

augmented buck converter.

REFERENCES

[1] Intel Corp., “Voltage regulator module (VRM) and enterprise voltage

regulator-down (EVRD) 11.1 design guidelines,