Definition
Plain language
A score saying how much less computing time one approach needed to reach the same quality as another.
As stated in the literature
Efficiency metric defined by taking the weaker run's final score and measuring when the stronger run first reached it; reported as a ratio of compute consumed to reach matched performance.
Also called: compute multipliers
Why it matters: It turns vague claims that one method is 'better' into a concrete statement about how much hardware time and money you save for the same result.
For example, if a weaker setup needed ten hours of processing to reach a certain accuracy and a better one reached that same accuracy in two hours, the compute multiplier is five.
Heard on the show
“Then there's the central measurement, the compute multiplier.”Episode 287 — Can You Measure Research Taste If The AI Isn't Allowed To Code?