PrimeShield:

PrimeShield is a new tool from Synopsys which can be run within PrimeTime Shell or can be run standalone. It does require a separate license for PrimeShield regardless of how you run it. It has it's own set of cmds, which will work, only once PrimeShield is invoked. We'll treat PrimeShield as an extension of PrimeTime.

PrimeShield tool performs Monte-Carlo-based design variation analysis. analysis is based on the results of PrimeTime parametric on-chip variation (POCV) analysis, which calculates the slack of a timing path as statistical
distributions rather than a single fixed values. By default, POCV analysis reports a path as meeting a timing constraint if the slack is at least zero at the -3σ (minus three sigma) point of the slack distribution, three standard deviations below the mean.

From a normal gauss distribution plot, we see that -3σ and +3σ points on the left and right sides of plot are the points within which 99.73% of the point lie. So, with slack on X-axis, only (100-99.73)/2 = 0.135% of the slacks will be less than -3σ slacks. So, if this was the only path on a die, then out of 10,000 dies, 13.5 dies would fail timing and would need to be thrown away. This will be our yield loss. We can improve the yield by making timing meet for 4σ, 5σ etc by changing the design. But that has power/area implication as larger and/or leakier gates will need to be added to meet more stringent timing. But on the plus side, we get higher yield.

How if we had many more paths on a die with all paths meeting > -3σ slacks? In such a case, for the chip to pass, all paths would need to pass. Assume there were 1000 such paths on chip, and all of them being independent. Then for the chip to pass, each path would would need to pass timing. Probability of that happening = (0.99865)^1000 = 0.26 , i.e only 26% of chips would pass timing for all paths. However, if we had 10K such paths, then prob = (0.99865)^10000 = 0.000001 or 0% yield. i.e only 1 out of 1M chips would pass timng for all paths.