The dataset behind every Interact recommendation rests on a single methodological foundation: SPEC SERT, the industry standard for server energy benchmarking. Every major server OEM (Dell, HPE, Cisco, IBM, Fujitsu) uses SERT to generate the Energy Star ratings shipped with their hardware. We use the same framework to train our machine learning model.
"Industry standard" is a defensible starting point. It is not, on its own, an argument. So we tested it.
What SERT does
SERT, developed by the Standard Performance Evaluation Corporation, measures server performance by running synthetic workloads against different subsystems and aggregating the result against measured power draw. The aggregation is weighted to reflect where energy actually goes in a typical server: CPU dominates at around 65% of the energy budget, memory at around 30%, storage at around 5%. The single output is an operations-per-Watt figure, comparable across hardware.
The strength of SERT is consistency. The same suite, run on different hardware, produces results that can be compared like for like. The criticism, common enough to address, is that any single framework is only as representative as the workload mix it assumes. The question is whether the answers SERT gives are comparable to those produced by other measurement routes.
Testing from first principles
We have run that test in the lab. SERT against the Phoronix Test Suite. SERT against the Intel Processor Diagnostic Tool. SERT against CPU torture tests designed to push hardware beyond sustained operational limits. All on the same representative server configuration, in a wind tunnel maintained within ASHRAE temperature and pressure guidance, with power measured directly at the PDU rather than via a SERT-approved analyser. The PDU approach removes any methodological favouritism and gives every framework a level playing field.
What we found
Three patterns emerged.
For conventional comparable workloads, SERT and the Phoronix alternatives produced energy draw within a percentage point of one another. The frameworks are measuring the same physical phenomenon by different routes and arriving at the same answer.
For CPU torture tests, the alternatives exceeded SERT by single to low double digit percentages. This is expected behaviour. Torture tests are designed to push hardware past sustained operating limits; the higher draw is a measurement of that overshoot, not a critique of SERT.
For the Intel Processor Diagnostic Tool, a similar pattern: peak draw above SERT, consistent with a tool designed to maximise hardware stress rather than simulate real workload behaviour.
The summary: SERT and well-known alternative benchmarks agree on operational workloads. The alternatives that exceed SERT do so for documented reasons consistent with their design.
Why this matters
Two things came out of the work that we did not have before it. First, independent confirmation that the benchmarking framework underneath our dataset is not an artefact of methodology. The same conclusions emerge from other measurement routes. Second, evidence that PDU-level direct measurement and the SERT-approved analyser methodology produce equivalent readings, which gives us a path to validate any third-party submission against an internal lab measurement.
For a customer relying on Interact's analysis, the practical implication is that the benchmark layer underneath every recommendation has been tested against first principles rather than taken on industry trust. The conclusions hold up regardless of which measurement route you arrive by.
That is what we mean when we say the dataset is defensible. The methodology behind it is, too.