PMBusTakes Command of Data Center Power Issues

Apr 1, 2008 12:00 PM
By Brian Griffith, Server Power Delivery Architect, Intel, DuPont, Wash.

Data Formats

The PMBus data format most preferred by system manufacturers is the linear format. This format provides the easiest method for system usage with good enough resolution for any of the power sensors mentioned previously. The PMBus linear format is as follows:

X = Y × 2^N , where Y equals an 11-bit two's complement integer (least significant bits), which is the mantissa; N equals a 5-bit two's complement integer (most significant bits), which is the exponent defining the scaling; and X equals a real-world value.

The system reads both Y and N to determine the real-world value of X. The N value is used to scale the real-world value and should be a constant for any given power supply. Using a constant value for N allows for easier calculations inside the power supply and the system.

The 11-bit Y value provides enough resolution for any system-usage model. Table 2 shows an example for an 800-W single 12-V output power supply with N scaling values and real-world value range.

Averaging and Accuracy

Accuracy of the PMBus sensors is critical. The power converters are becoming a more integral part of the system, and poor accuracy can have a negative impact on system performance. Because the system operates over a 20% to 100% load range, depending on the system operating state, accuracy over this entire range is important. Understanding how to sense ac input current and voltage to determine ac input wattage is needed before starting the design.

Sampling the 50/60-Hz ac voltage and current at a high enough frequency to account for distortion is important. All server ac-dc power supplies have active power factor correction (PFC); therefore, this sampling rate can be limited to approximately 5 kHz.

Each PFC implementation is different in how well it limits current harmonic, so each design should measure this and set the appropriate current sampling rate. The loading condition where it is hardest to achieve good accuracy is at high input voltage (240 Vac) and at light loads (20% of maximum). This is the point where the power factor is the worst and where the input current sensor has the lowest signal-to-noise ratio.

The goal of most system manufacturers is to achieve better than ±5% accuracy over 20% to 100% of the converters' operating range. If the best accuracy is required, some manufacturing line calibration may be needed.

Because most server computer systems will poll the converters at a rate equal to about once per second, considerations must be taken to prevent aliasing of the data. The power converter must act as the anti-aliasing filter by averaging the data inside the power supply to prevent error due to slow system-polling rates.

The period of this averaging must be 1 second to prevent aliasing. This creates a challenge for the power converter. Increasing the system polling rates of the power converters can help to alleviate this challenge. However, it will create its own challenges in the server management design. There are basically three options for averaging sensor data inside the power supply:

• Infinite impulse response (IIR) filter

  The IIR filter can be implemented with a standard first-order difference equation. Difficulties arise due to the large difference between the sampling rate and the bandwidth. The sampling rate is more than 1000 times the filter bandwidth. This creates coefficients in the difference equation very close to 1 and 0, which requires more than a single-byte calculation method. However, this can be managed with the proper coding. The advantage of this method is that it generates a true running average of the sampled data.

• Arithmetic average

  By simply summing the sensor data over the averaging period and dividing by the averaging period, the average over that period can be determined. Then repeating this for the next period will generate another average value and so on. This will not give a true running average and, therefore, will add some amount of error if the load is very dynamic. The advantage of this method is its simplicity.

• Running arithmetic average

  By saving all the sensor data points over the average period, newer data can be added to the average and the old data can be dropped from the average. The problem with this method is the need to save large quantities of data: 1000 data points if sampling at 1 kHz. The advantage is the accuracy of a true running average combined with an arithmetic average.

Isolation of Sensors

Since some of the power sensors are located on the primary side of the ac-dc power supply, these ac current, voltage and power sensors must pass information to the secondary-side PMBus microcontroller and still meet safety regulator requirements. Depending on which method is used to average the data, the speed at which the primary-side sensor data is communicated to the secondary side may be critical.

Standard optocoupler devices used to isolate primary- from secondary-side signals in a power supply are too slow for any of the schemes requiring secondary-side calculations. There are three methods for implementing primary-side sensors:

• Primary-side calculation

  In this method the input voltage and current are sampled with a primary-side referenced microcontroller. This same microcontroller processes the data to generate the root-mean-square (rms) voltage, rms current and power in watts. The primary-side microcontroller then communicates the resulting data to the secondary. This means there is no need for a fast primary-to-secondary-side communication path. The drawback is the need for a primary-side microcontroller that can perform the needed calculations.

• Secondary-side sensing and calculation

  In this method, there is no primary-side microcontroller. The primary voltage and current are either passed to the secondary side via isolated sensors, current-sensor transformers or pulse-width-modulation signals over optocouplers. This eliminates the primary-side microcontroller; however, it requires more complicated isolated sensing circuits.

• Primary-side sampling/secondary-side calculation

  In this method, the primary-side microcontroller mainly serves as the analog-to-digital converter. The raw data is then passed to the primary side for averaging. This keeps the primary-side microcontroller simple; however, it requires a fast, isolated communication method.

Acceptable Use Policy blog comments powered by Disqus