Current Sensing For Server Power Monitoring: MOSFET Or Shunt?
Find a downloadable version of this story in pdf format at the end of the story.
Accuracy is important in server power monitoring. Inaccuracy in power measurement costs the system manufacturer in power supply, thermal, and other hardware and performance overheads. When linked to power capping systems, inaccuracy leads to computing power being “left on the ”. Sources (1,2) have established the cost of power in the server at $2.74. For a 1000-watt server, every 1% of power measurement inaccuracy is a 10-watt loss worth $27.40.MOSFET CURRENT SENSING DESIGN
MOSFET current sensing technology depends on the close matching of transistor cells within MOSFETs. A MOSFET is composed of thousands of transistor cells in parallel. These cells are identical as are their currents, which combine to become the drain current. It is possible to isolate the source connections of several cells and bring them out onto a separate sense pin.
This arrangement is not too unlike any classic MOSFET current mirror circuit with proportionately sized elements. Like any mirror, the closer the operating parameters of the two devices are, the more accurate the current ratio. And as any semiconductor engineer understands, the transistor-transistor current mirror is never a precision device when required to operate over wide variety of operating points, conditions, and temperatures. Fig. 1 depicts the simple model of the current sensing MOSFET.
The variety of methods to take advantage of MOSFET current sensing is well described in manufacturers literature and application notes (see appendix following feature). For our purposes, we'll consider only the most accurate implementation, sometimes called the “virtual earth” method, that uses an amplifier to force the same drain to source voltage across both the sense and main element. The basic virtual earth is shown in Fig. 2.
As will be discussed in more depth later, difficulty arises with this accurate circuit at high currents. A common current sensing ratio is 500:1, and in a 50A circuit, the sense current will be as high as 100mA, more than many op-amps will happily provide. The solution to this problem will also solve another difficulty encountered with the circuit of Fig. 2, that the output is referred to the Kelvin source and not to ground, and the complete version of this circuit requires an additional difference amplifier to deliver a ground referred output. If the source is grounded, Fig. 2 will also require a negative power supply for the op-amp, and the op-amp output will swing negative. The solution shown in Fig. 3 adds a transistor to “cascode” this current into ground rather than an op-amp output, providing an accurate ground referred output along with the ability to handle higher currents.
There is a constraint on the circuit of Fig. 3 in that the source of the Q1 must be at a sufficient positive voltage to provide sufficient working voltage for Q2. For example, if up to 3 volts is required at the drain of the Q2, then the Q1's source cannot be lower than 3 V, in addition to whatever drop the Q2 requires. This constraint is no problem in server hot swaps where the source will always be at 12 V, although this circuit, much like its companion used for sensing shunts, suffers from the difficulty that a complete short of the measured line results in zero output. In simple terms, you cannot measure current during a complete short circuit event, without adding considerable additional circuitry.
Now that the essentials of the most precise circuit methods have been shown, some practical performance capabilities can be discussed.
CURRENT RATIO
The most fundamental consideration is current ratio, a parameter always provided on the data sheet. As will be shown, higher is better. When reviewing the history of current sensing MOSFETs it is evident that it is difficult to find ratios above 500 at this time. In the past devices were offered with ratios as high as 2600 but have since been discontinued (International Rectifier seems to have completed eliminated offerings in this space). Customary engineering practice also suggests that higher ratios would not be the panacea they seem at first glance. It is well known that linearity for example, can suffer greatly at high ratios. From a practical standpoint, you select a current sensing MOSFET that meets your voltage, current, and SOA requirements and accept what you get for current ratio.
For example, a 24V 10A device with a ratio of 250, 25V 149A device with a ratio of 404, and a 40v 155A device with a ratio of 500 are examples used several times in this paper. Looking at readily available choices shows the most common ratio you'll encounter is around 500.
If you are doing very high current applications, it is essential to remember just how much current you have flowing in the sense transistor. A 100-amp application will have to deal with 200mA from this sense output.
CURRENT RATIO ACCURACY
Accuracy is defined by the current mirror ratio, n. The virtual ground circuit discussed here contributes inaccuracy due to amplifier offset and the tolerance of the added resistor. But for the sake of discussion, consider that an integrated IC and MOSFET approach with trimming will reduce resistor tolerance to negligible values by being calibrated at 25°C and at the specified maximum current point. So the relevant errors then become linearity and temperature coefficient.
Again, studying some example data sheets reveals the following specifications (Table 1), unless noted, MIN/MAX specifications are referred to. This selection includes the lowest R( DS ON) current sensing MOSFET known at this time:
Again, considering that an integrated approach could trim room temperature errors, remaining errors in published specifications are due to tempeature coefficient, but additional errors will be encountered over the range of current and gate drive. Note that the ratio varies over current, indicating non-linearity, and further examination of data sheets shows a dependency on the exact amount of gate drive. The graph of Fig. 4 is typical of current sensing MOSFET performance. Even considering a limited range, for example 8 to 12 V of gate drive, shows a variation of ±3%. To put this into perspective, a 1.3 V variation in 10 V gate drive introduces 1% of error. Achieving maximum accuracy indicates the need to control gate drive within ±0.13 volts for less than 0.1% contribution to error. This means the gate drive alone must be controlled to around ±1%. For error considerations, we will make the assumption that any integrated approach implements tight gate voltage control.
A big problem is the lack of MIN/MAX limits on temperature coefficient from any manufacturer, typical values that range from +/-0.006%/°C to as high as ±0.078%°C make estimating maximum values difficult. Parameters that do not have guarantees indicate that they are poorly controlled, or difficult or uneconomical to control. Credible literature (3) indicates that around ±0.11%/°C is more representative of achievable performance and this suggests a trimmed integrated approach could be specified at ±7%, only at full scale, from 10°C to 85°C. In order to put MOSFET sensing in its best light with currently available parts, with difficulty it may be possible to achieve ±3% over 10°C to 85°C at full scale.
If measurement is expected over a range of 5:1 or more of current an additional 9% nonlinearity needs to be included. At its best, MOSFET current sensing over a range of current is a ±16% proposition assuming all other stabilizing methods discussed here are implemented (e.g., tight gate voltage control).
EFFICIENCY
Efficiency is most often mentioned as the key reason to use MOSFET current sensing. Articles have been published even describing it as “lossless”(4)(5). It is useful to closely examine the efficiency of this method. Again, sense ratio plays a key part in efficiency. As an example let's use a 12V 50A server with MOSFET sensing as shown in Fig. 5. It is readily obvious that at 50A there is 100mA of sense current, 1.2 W of power lost in measurement. Consider the effect on the circuit output load resistor. A typical range for an ADC input might be 3 V full scale, in which case 300mW needs to be dissipated in the resistor, heating it and causing resistor temperature efficient errors. Or, looked at in another way, the load resistor needs to have a good temperature efficient, increasing its cost.
This also needs to be considered in light of the fact that most recent shunt sensing ICs offer a full-scale shunt input of 25mV. In the past, many designers have routinely considered 100mV the typical drop required on a shunt, and 50mV has been sort of a “de facto” standard in industrial systems.
Certainly a higher sense ratio can improve efficiency, but as this is written sense ratios above 1000 exist only on MOSFETs with high RDS(ON). It can be expected that accuracy will be even more difficult to achieve at those higher sense ratios. At the outset of this article the cost of power in the server was mentioned. Evaluating the value of MOSFET current sensing needs to take in not just efficiency, but accuracy and component cost as an overall assessment of value.
COMPARING MOSFET WITH SHUNT
In this case, we show a 12 V, 50A MOSFET current sensing system capable of:
-
500:1 sense ratio
-
±3% accuracy at full scale from 10°C to 85°C
-
±12% over a 5:1 dynamic range
-
1.2 W of power lost.
For comparison, let's consider the state of the art in shunt resistor sensing based power monitoring as shown in the schematic of the National LM25066 in Fig. 6. These shunt-sensing devices require a 25mV drop across the shunt at full current, so at 50A the loss in the shunt is 1.25 W. The power monitor IC alone (not including the 1% shunt) specifies±1% accuracy at full scale:
-
±2% (includes shunt) at full scale from 10°C to 85°C
-
±6% at 20% of full-scale range (includes shunt, 5:1 dynamic range)
LOSSLESS?
The MOSFET steals a little current, the shunt steals a little voltage. Neither is lossless, and state of the art shunt implementations are more than competitive. In servers, the efficiency of the current measurement method must be considered along with accuracy to determine the value of the approach.
REFERENCES
Server Power Cost, Power Measurement Accuracy and Efficiency:
- From: http://blogs.amd.com/work/2010/07/06/one-watt-is-a-lot/ “One Watt is a Lot,” July 6 2010 Anita Tulsiani of AMD
- Source: http://www.liebert.com/common/ViewDocument.aspx?id=880. 0.00284KW (2.84 W converted to Kilowatts) × 8760 (annual hours of server operation) × $0.11 (assuming KwH rate for energy usage) = $2.74/watt From: “Energy Logic: Reducing Server Energy Consumption by Creating Savings that Cascade Across Systems” Emerson Network Power SL-24621(0809) Consequently, every Watt of savings that can be achieved on the processor level creates approximately 2.84 W of savings for the facility
- Current-sensing power MOSFETs with excellent temperature characteristics Hidefumi Takaya, Kyosuke Miyagi, Kimimori Hamada TOYOTA MOTOR CORPORATION, 543 Kirigahora, Nishihirose-cho, Toyota, Aichi, 470-0309, Japan Tel: +81-565-46-3377 Fax: +81-565-46-3382 E-mail: takaya
- NXP AN10322: In 2004, section 5, the Conclusion of AN10322 was worded: “SenseFET devices are an effective means of protecting automotive applications. They are a loss less and cost effective alternative to traditional current shunts whilst retaining realistic tolerances.” Revised by 2009 to say: “SenseFET devices are an effective means of protecting automotive applications. They are a low loss and cost-effective alternative to traditional current shunts whilst retaining realistic tolerances.”
- ON Semiconductor AN8093/D: “When a signal level resistor is connected between mirror and source terminals, a known fraction of load current is sampled without the insertion loss that is associated with power sense resistors. For this reason, the technique of measuring
ADDITIONAL REFERENCES
- Current Sensing Trench Power MOSFET for
Automotive Applications
Y. Xiao, J. Cao, J. D. Chen and K. Spring
International Rectifier Corporation
222 Kansas Street, El Segundo, CA 90245, U.S.A.
Tel.: 310 726 8792, Fax: 310 726 8846, e-niail: [email protected]
Describes development of high ratio sense MOSFET. - Power MOSFETS Theory and Applications, Chapter 19
Duncan A Grant, John Gower
Wiley-Interscience 1989-04 - Manufacturer Application Notes:
NXP AN10322 “Current sensing power MOSFETS” - On Semi AND8093/D “Current Sensing Power MOSFETs”
- On Semi AND8210/D “Current Sensing Power MOSFET Use in DC-DC Converters”
- Vishay AN606 “Current-Sensing Power MOSFETs”
- International Rectifier AN959 “An Introduction to the HEXSense Current- Sensing Device”
Download the story in pdf format here.