Improving Energy Efficiency in AI Data Centers with Power Integrity Testing

The following chart illustrates the impact of the growth in energy demand driven by the deployment of artificial intelligence (AI)-based technologies in data centers and processor-intensive servers. The International Energy Agency (IEA) predicts that by 2030, data centers will account for 7% of global electricity consumption, equivalent to the electricity consumption of India.

Figure 1: Power consumption of data center CPU and GPU technologies.

As power demands continue to grow, focusing on energy efficiency is critical. Tektronix has partnered with renowned power integrity expert Steve Sandler to develop superior measurement techniques to improve operational efficiency/energy efficiency in next generation AI data centers.

Improving the energy efficiency of power distribution networks (PDNs) Improving the energy efficiency of power distribution networks (PDNs)

The PDN must provide many low-noise DC power rails to drive sensitive GPU loads in server racks. The drive for higher speeds and higher densities means faster edge rates, higher frequencies, and more rails at lower voltage levels and higher currents. This highlights the importance of good power integrity.

The purpose of power integrity testing is to verify that the voltage and current reaching the point of load (POL) meet the load's power rail specifications under all expected operating conditions. Accurately measuring millivolt-level power rail noise at gigahertz frequencies is particularly important.

Let's look at how to evaluate PDN performance by looking at a high-level diagram of the power distribution network in a server system.

Figure 2: High-level power distribution network in a data center.

As shown in the figure, a typical data center powers its AI-based servers from a 12 V, 24 V, or 48 V DC power supply, which then converts the voltage to other supply voltages on the motherboard. With visibility into every link in the chain from the power supply output to FPGAs, processors, and other complex ICs, engineers can control the power rail impedance to very low levels to deliver the high currents required by AI servers driven by GPU technology. Impedance management is tricky because the power distribution network is made up of many impedances, including voltage regulators, decoupling capacitors, and PCB traces. High-speed switching and hot-swapping server cards can introduce unexpected impedance changes, which can cause excessive transients or noise.

Minimizing noise in the PDN is the first step to ensuring a stable and energy-efficient design. Power rail noise specifications can range in frequency from hundreds of megahertz to several gigahertz with amplitudes in the millivolt range.

Evaluating energy efficiency begins with power quality measurements of the AC line input and output to ensure that the line voltage and line current meet requirements. The measurements used to evaluate quality are as follows:

• frequency

• Effective voltage and current

• Effective value

• Impedance

• Crest Factor (Voltage and Current) Crest Factor (Voltage and Current)

• Active power, reactive power and apparent power Real power, reactive power and apparent power

• Power Factor and Phase Power Factor and Phase

To ensure these measurements are made accurately, the choice of oscilloscope probes is important; use a differential probe to measure the system's line voltage and a current probe to measure the system's line current.

Another key measurement is the frequency response analysis of the PDN control loop response. This will provide important information about the control loop speed and the stability of the power supply. The analysis results are viewed with the help of a Bode plot, with an example setup in Figure 3.

Figure 3: Measurement setup for power distribution network impedance.

Power integrity detection system deserves attention Power integrity detection system deserves attention

The high impedance 10X passive probes that come with today's oscilloscopes may have adequate bandwidth, but will attenuate the noise signals you want to measure. A 1X probe will pass the noise signal without attenuation, but its bandwidth is only a few hundred megahertz. A transmission line probe or cable with 50Ω input impedance has excellent high frequency performance, but will create significant loading at DC unless a DC isolator is added. An attenuating transmission line probe creates less loading while maintaining low noise and high bandwidth.

Power rail probes are another type of low noise probe with an offset range up to 4 GHz and a DC offset range of -60 to +60 Vdc. Power rail probes are a more accurate alternative to traditional passive probes for identifying noise sources, as shown in Figure 4 below. Depending on the voltage of the power rail, a DC block may be required. If required, ensure that the DC block provides surge protection for the oscilloscope and is not affected by DC or AC bias. Power rail probes, while capable of measuring very low noise, are also single-ended measurements. Therefore, a coaxial isolator is required that can further reduce measurement ground loop errors. Picotest offers a variety of DC blocks and coaxial isolators to meet this need. Learn more about the ultimate power rail noise measurement.

Figure 4: Comparison of power line ripple measurements using a passive probe (lower trace) and a power rail probe (upper trace).

Fast, low-noise acquisition combined with ultra-fast edge loads simulates AI-class processor workloads, allowing accurate evaluation of power rail noise voltages and rail-to-rail crosstalk in PDN designs. When used in conjunction with a Tektronix 5 Series B MSO or 6 Series B MSO oscilloscope; Picotest offers a complete line of load devices up to 2,000 amps, 1 nanosecond edge loads, and supports sampling rates up to 65Ms/s for accurate simulation experiments. (See Figure 5) Fast, low-noise acquisition combined with ultra-fast edge loads simulates AI-class processor workloads; allowing accurate evaluation of power rail noise voltages and rail-to-rail crosstalk in PDN designs. When used in conjunction with a Tektronix 5 Series B MSO or 6 Series B MSO oscilloscope; Picotest offers a complete line of loads up to 2,000 amps, 1ns edge loads, and supports sampling rates up to 65MS/s for accurate simulation work. (See Figure 5)

Figure 5 shows the characterization of an AI-class processor subjected to a pseudo-random high-amplitude load.

Characterization using Picotest load equipment and measurements with a Tektronix 6 Series B MSO oscilloscope ensures accurate characterization. The Tektronix 6 Series B MSO oscilloscope is the ideal instrument for capturing low noise, high resolution signals. Oscilloscope measurement analysis helps save time and reduce errors

Identifying and analyzing trouble spots in a PDN can be time consuming. Finding ripple, overshoot, undershoot, turn-on, turn-off, time trends, settling time, and jitter signals in a power distribution network is a complex task. Thankfully, most modern oscilloscopes today offer built-in analysis software for setting up the instrument and automating signal acquisition and display. Below is an example of an automated ripple measurement. Having these features built into the instrument, along with the ability to automate via a remote PC, can simplify AI performance evaluation for large teams, and it is also possible to evaluate how AI-enabled performance changes over time and temperature to test server efficiency and durability.

Figure 6: Automated ripple measurement with annotated results shown on the right side of the 5 Series B MSO oscilloscope display.

Summarize

As artificial intelligence (AI) drives increased energy demands in next-generation data centers, evaluating the performance and efficiency of power distribution networks (PDNs) has become more important than ever.