SimPoint is the industry standard for CPU simulation. Instead of running a full application — billions of instructions — it selects representative code regions and simulates only those. The assumption is that a few well-chosen samples capture the whole. Ekman (arXiv:2603.22605) shows the assumption fails: with 20 SimPoints, some applications show 40–60% prediction error.
The fix is not more SimPoints. It's better sampling.
Ekman decomposes SimPoint into two operations: stratification (clustering code regions by similarity) and selection (choosing which regions to simulate within each cluster). SimPoint does both simultaneously — it clusters and picks centroids in one step. The two-phase approach separates them. First, take a large preliminary random sample to estimate the variance structure. Then use that structure to allocate simulation budget across strata proportionally to their variability. The result: maximum per-application error drops to 3%, with an order-of-magnitude reduction in required sample size.
The key insight is that SimPoint's clustering step, while identifying natural strata, doesn't allocate samples within strata according to their variance. A cluster with high internal variability gets the same number of samples as one with low variability. Stratified sampling corrects this — it sends more samples where the uncertainty is highest. The budget goes where it's needed, not where it's convenient.
This is a textbook statistics result applied to a domain that reinvented sampling without consulting the textbook. The two-phase stratified estimator has known variance reduction properties and confidence interval formulas. SimPoint has heuristics and hopes. The gap between 40–60% error and 3% error is the gap between ad hoc engineering and statistical theory.
The broader lesson: when a simulation tool becomes standard, its limitations become invisible. Everyone uses 20 SimPoints because everyone uses 20 SimPoints. The error isn't measured because the tool doesn't provide error bars. Ekman's contribution is not a new simulation technique but the application of a known sampling framework to a domain that needed it decades ago.