Model to silicon correlation viewer

How it works

Before a chip ships, the models that signed it off are compared with what the silicon does. This viewer takes two pasted tables, measured results and simulated or characterised values, joins them by parameter and condition, and shows the difference, the percent difference, how far the model sits from the measured spread, and which points break the limit you set. A fit of measured against simulated for each parameter shows whether a miss is an offset, a scale error or noise.

1. Join the tables. Each row is a parameter, a condition (process corner, voltage, temperature, or any label), a value and a unit. Rows with the same parameter and condition are matched, ignoring case and spacing. Several measured rows for one pair are several units or dies: the tool takes their mean and sample standard deviation. Units with an SI prefix are converted to the simulated unit (mV to V, uA to nA).mean = sum(x) / n; s = sqrt(sum((x − mean)^2) / (n − 1))
2. Difference and percent difference. The difference is the measured mean minus the simulated value. The percent difference divides it by the simulated value, or by the measured mean if you choose that reference.difference = measured mean − simulated; percent = 100 x difference / |reference|
3. z-score against the measured spread. With three or more measured units, the z-score says how many standard deviations the model value sits from the measured mean. A large z means the model is outside what the silicon population shows, whatever the percentage. It is shown beside the limit check, not instead of it.z = (simulated − measured mean) / s
4. Limits. Each parameter has an allowed difference, either per cent of the reference or an absolute figure in the parameter's unit. A point is flagged when the size of its difference is above the limit. Parameters without a limit of their own use the default.flag when |percent| > limit (per cent), or |difference| > limit (absolute)
5. Correlation and fit across conditions. For each parameter the tool fits the measured means against the simulated values by least squares: measured = slope x simulated + offset. A slope near 1 and an offset near 0 is a model that tracks silicon. A slope away from 1 is a scale error. A good correlation coefficient with a large offset or slope error means the model follows the trend but is biased, which is a calibration problem. A poor coefficient means the model misses the trend. The coefficient needs three or more conditions.slope = Sxy / Sxx; offset = mean(y) − slope x mean(x); r = Sxy / sqrt(Sxx x Syy)
6. What-if on the limits. The limit table recounts the flagged points with every limit scaled by 0.5 to 2, to show how many points are close to the line.

Sources

The methods and default figures rest on these public sources. The wording here is our own; read the source for its full text.

• Model to Hardware Correlation, Silicon Lifecycle Management use case datasheet (opens in a new tab), Synopsys, Undated vendor datasheet, read October 2026. Why pre-silicon models and silicon drift apart, and the idea of checking corner targets against measured ring oscillator results.
• Semiconductor Measurement Technology: Evaluating a Chip, Wafer, or Lot Using SUXES, SPICE, and STAT2 (NIST Special Publication 400-90) (opens in a new tab), J. C. Marshall and R. L. Mattis, National Institute of Standards and Technology, April 1992. An early method for comparing simulated and measured test chip results, with a ring oscillator example.
• NIST/SEMATECH e-Handbook of Statistical Methods, section 4.1.4.1, Linear Least Squares Regression (opens in a new tab), National Institute of Standards and Technology, Online handbook, current edition. The least squares line and the correlation coefficient used for the fit across conditions.