With the exponential increase in output database size due to the aggressive optical proximity correction (OPC) and resolution enhancement technique (RET) required for deep sub-wavelength process nodes, the CPU time required for mask tape-out continues to increase significantly. For integrated device manufacturers (IDMs), this can impact the time-to-market for their products where even a few days delay could have a huge commercial impact and loss of market window opportunity. For foundries, a shorter turnaround time provides a competitive advantage in their demanding market, too slow could mean customers looking elsewhere for these services; while a fast turnaround may even command a higher price. With FAB turnaround of a mature, plain-vanilla CMOS process of around 20-30 days, a delay of several days in mask tapeout would contribute a significant fraction to the total time to deliver prototypes.
Unlike silicon processing, masks tape-out time can be decreased by simply purchasing extra computing resources and software licenses. Mask tape-out groups are taking advantage of the ever-decreasing hardware cost and increasing power of commodity processors. The significant distributability inherent in some commercial Mask Synthesis software can be leveraged to address this critical business issue.
Different implementations have different fractions of the code that cannot be parallelized and this affects the efficiency with which it scales, as is described by Amdahl’s law. Very few are efficient enough to allow the effective use of 1000’s of processors, enabling run times to drop from days to only minutes. What follows is a cost aware methodology to quantify the scalability of this class of software, and thus act as a guide to estimating the optimal investment in terms of hardware and software licenses.