Passmark Inline PSU Tester - Excessive CPU Ripple

[zmem]

I've recently been testing power supplies in a server environment that has dealt with abnormally high numbers of CPU failures (>6%). My test is to run the Inline PSU Tester while the server performs its typical workload (~30-50% load on average, with large fluctuations). It seems every PSU I've tested, part of this environment or otherwise, always seems to report suspiciously high ripple on the 12V CPU rail in this test, as well as the 12V, 3.3v, and 5v rails on some units. Here are some ripple averages over several minutes as an example:

Minimum PSU wattage: 220w
Average power draw for all samples: ~120w

Sample #1 (Experienced CPU failure): 12v: 94.5mv; 5v: 51mv; 3.3v: 24.19mv; 12v CPU: 191.95mv
Sample #2 (Brand New): 12v: 146.1mv; 5v: 59.27mv; 3.3v: 34.7mv; 12v CPU: 179.87mv
Sample #3 (Lightly used, no CPU issue): 12v: 50.7mv; 5v: 34.76mv; 3.3v: 35.77mv; 12v CPU: 163.11mv
Sample #4 (Quality Desktop PSU): 12v: 66mv; 5v: 40.76mv; 3.3v: 28.33mv; 12v CPU: 115.38mv

I can post the full logs if preferred, but this is the data that stands out to me. Even #4 reports, what I consider, high CPU ripple, albeit within the ATX specification. Every other unit appears to be dangerously high, although they are budget class units. Is there anything in my testing methodology that could cause such a high number? Does 192mv average even make sense? That number almost seems too high be accurate.

I should also note that Sample #2 and all of the identical models (that I've tested) report voltage sequencing errors, even when not set to strict mode.

[David (PassMark)]

The specification of ripple measurement is vague in the ATX standard.

The standard says: Ripple measurement shall be made with an oscilloscope with 20MHz bandwidth. Outputs should be bypassed at the connector with a 0.1uF ceramic disk capacitor and a 10uF electrolytic capacitor to simulate system loading. Our PSU tester follows this requirement.

We are measuring against a common ground (single ended measurement). Differential measurement can give slightly different results, but a lot more internal channels are needed for this (i.e much more expensive gear). The standard is vague on what should be the correct method.

The way we measure the ripple is as follows. We sample the voltage on each rail at 10K samples / second. We then take 1 sec of data and search for the absolute min and max of each rail, every second. The min/max difference is reported as ripple. The standard doesn't include any details about the measurement time period. But the measured value depends a lot on what time window is used and the sample rate. The voltage measurement precision in our tester is +-4 mV, due to quantization error in the A/DC. So numbers around the 4mV level are also not going to be accurate.

Importantly, it is common practice to measure ripple under a constant steady load. Could be high load or low load, but it should be constant. Having the machines run their normal variable load is sure to increase the ripple measurements and also make them less consistent. For example, a sudden increase in system load will naturally cause a voltage drop and ringing until the system stabilizes but I don’t believe the specification would classify that behavior as ripple.

[zmem]

Importantly, it is common practice to measure ripple under a constant steady load. Having the machines run their normal variable load is sure to increase the ripple measurements and also make them less consistent.

For example, a sudden increase in system load will naturally cause a voltage drop and ringing until the system stabilizes but I don’t believe the specification would classify that behavior as ripple.

Thank you for your response. Could the data I have be indicative of another issue, such as poor transient response? Although the system load is variable, the most questionable power supplies have been reporting the highest amount of ripple, well outside of the ATX specification.

During a consistent load, even the questionable power supplies stay mostly within spec, though I would consider them borderline.

[David (PassMark)]

Without wanting to generalize too much, lower ripple is associated with better power supplies.
I don't have any actual data, but I would not be surprised if the units with higher ripple also had poor transient response.