Comparing hot swap IC solutions in server power reporting (Part 1 of 2)
Abstract
High-performance servers have long used a power-measurement method that is based on analog-current sensing, filtering, and calibration circuits providing signals to an ADC contained in the Baseboard Management Controller(BMC).This system typically uses a calibration circuit to provide state of the art accuracy, although the shunt resistor itself is not included within the calibration loop. The component count and board real-estate requirements of this circuit are high.
Power density is not the only issue in today’s data center; operating cost has been another hot topic as the number of servers rose in the past years. Many techniques such as power capping have been introduced to increase energy efficiency in data center, but the ability to measure the power level accurately remain the key factor. This paper compares the area, cost, power saving, and other benefits gained with an integrated solution used in protection and power measurement in servers compared to discrete implementations.
Introduction
The significance of error in power measurement in servers is well known. Recent published studies have estimated a cascade effect such that for every watt in savings at the server component level, 2.84 watts can be saved from the facility energy consumption (see Reference). For example, consider a 600 W server where a 5% error represents 30 W. This represents an unfavorable 85.2 W impact in facility energy consumption. In a typical data center with 1,000 servers (Reference), this adds up to 85.2 kW of wasted energy. By reducing the power accuracy error to 2%, 34 kW of energy is wasted – a 60% reduction.
With the numbers of servers rising in data centers, the financial impact of the inaccuracy can quickly add up to millions of dollars in extra utility bill. It is estimated that the annual cost in operating data centers in the U.S. has reached as high as $3.3 billion (Reference).
There are number of considerations in the power management error, the primary one being errors in power capping levels that could deprive the end user of full levels of computing power. Other considerations are necessary system power and cooling margins.
Servers continue to grow in power and capability, and as they do they consume more circuit board real estate with the core functions of the server, reducing the room left for necessary support circuitry including power management and monitoring. There is no question about the necessity of monitoring.
In addition to the financial benefits already discussed, the simple fact is that end users are demanding power consumption information. There is a need to make these power-monitoring systems take less room inside the server, yet with no performance loss.
With those considerations in mind, the following is a summary of key benefits of an integrated solution compared to discrete implementation including:
1. Board area and cost
2. Power calculation and accuracy
3. System architectural impact
4. Advanced protection/reliability capabilities
It should be noted that B costs are materials only. Labor, pick and place, and associated assembly costs are not included. Components that are identical in all approaches are not included, for example the shunt resistor is required in all approaches and takes up the same board real estate and cost.
1. Board area and cost comparison
A simplified schematic comparison of all three approaches is shown in Figure 1 and Table 1.
Figure 1: Simplified schematic comparison of all three approaches.
Note that board area does not include components shared
by all three methods including shunt and MOSFET.
Table 1
2. Power accuracy and calculation comparison
While the legacy systems can capture voltage and current in a single command, processing is still necessary to provide power calculation, averaging, setpoint actions. Even then, the quality of the averaging is directly related to how continuous and frequent the data readings are.
By providing on-board power calculation and averaging, the LM25066 makes it possible to read data "at the systems leisure", without the overhead of calculations. The LM25066 also provides a watchdog setpoint for power which can be used to trigger power capping. All of this functionality is provided along with the same level of accuracy as is achieved with the existing system.
Current Measurement
An essential element of data gathering to measure server power is the measurement of current into the server. Current measurement is not a trivial instrumentation exercise, since this must be done via high-side shunt-current sensing. These high-current systems cannot have their grounds disturbed in the way that a shunt on the low side would.
It is tempting to consider the various magnetic sensing methods available, but state of the art accuracy demands are now tighter than 3% for the entire measurement system including the shunt, and have dynamic range requirements that are in excess of 5 to 1 (meaning good accuracy is necessary down to a point at 20% of the input power).
To achieve 3% overall accuracy in the power measurement means that each channel–the voltage measurement and the current measurement channel–must have less than 1% error. The remaining 1% will come from the use of commonly available 1% tolerance shunt resistors for current sensing. High-side sensing with such accuracy demands very high common-mode rejection capability on the current sense.
There are the following considerations in computing total error:
1. The method for totaling errors. In uncorrelated errors it is customary to use root-sum-square (RSS) methods. Some designers have a personal preference for a straight sum, which is certainly a demanding, worst-case scenario. This paper will use root-sum-square in any uncorrelated error summaries.
2. The error in the voltage channel and the current channel multiply each other to determine the power error. At small error values, generally less than 3% total, the multiplied value is close to a simple summation of the errors.
To compute total error, first take the total error of the current channel, which has shunt error as well as instrumentation error, Equation 1:
Where:
RSERR = Shunt resistor tolerance
IIERR = Current channel instrumentation tolerance
As an example, a system using a 1% shunt with an instrumentation system specified for 1% error would have a total of 1.4% error in the current measurement channel.
The total error of the power measurement is provided by Equation 2:
Where:
VERR = Voltage channel instrumentation tolerance
PERR = Total error of power measurement
A system with 1.4% error in the current channel, along with 1% error in the voltage channel, will have an overall error of 2.4% in the power measurement, very close to the sum of the total errors of the two channels.
As has been shown, providing acceptable levels of total overall accuracy require at least 1% accuracy on the voltage and current channels of the instrumentation, to yield this overall 3% or better power accuracy.
Figure 2: Server instrumentation for power monitoring has high demands for accuracy..
Designing for a 3% overall power measurement will require better
than 1% accuracy on each channel of the power and voltage measurement.
Specific current-sensing considerations
It is often tempting to consider magnetic current-sensing methods at these currents, along with the high-side requirement. However, current state-of-art magnetic sensing method can only achieve current-measurement accuracy of 5%. Current sensing using shunt resistors still provides the best state-of-art in achieving 1% current-accuracy measurement.
While lower shunt drops reduce losses, generate less heat, enable smaller shunts, and take less real estate, the accuracy in current sensing always improves with larger shunt drops. The trick is optimizing the tradeoff of shunt loss against the necessity for high accuracy.
Voltage measurement
The voltage sensing is straightforward and as simple as a voltage divider into an analog-to-digital converter. However, the voltage divider involves two resistors and their effect on accuracy, and details such as this have made integrated solutions attractive. The voltage channel must be accurate because of variations in voltage. It is tempting to assume the voltage in a server would be somewhat constant, but consider that only 120 mV of variation corresponds to 1%.
Once satisfactory solutions have been arrived for current sensing and voltage dividers, there is still the matter of analog-to-digital conversion (ADC). Many choices are available for ADCs, many of which are integrated into the processors that would be used for a BMC.
However, these ADC’s performance such as offset, gain error, as well was linearity are not good as it appear. Those parameters, or at least gain parameters, are directly influenced by the ADC reference. References built-in the processors are invariably low performance references, unable to provide the accuracy required even over limited temperature ranges unless calibration is used. Unlike ADC that is integrated into a hot swap with power measurement device such as LM25066, one must take the additional ADC error into consideration.
(End of Part 1; Part 2 will look at an integrated hot swap and power measurement solution in detail and its comparative attributes.)
Reference
•Energy Logic: Reducing Data Center Energy Consumption by Creating Savings that Cascade Across Systems. Emerson Network Power, 2009.
About the authors
Joy Taylor is a product marketing manager at National Semiconductor Corp., focused on power management products and technologies. Taylor has held various roles in applications and product marketing at National since 2003. She has a BSEE from San Jose State University and an MBA from Santa Clara University.
Jerry Steele is a strategic applications engineer at National Semiconductor’s Tucson Design Center, specializing in defining power management products. Steele has over 25 years of experience in the analog and mixed-signal industry, has authored several articles and co-authored four patents.
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
