The problem is real, said Gary Smith, chief analyst at Gary Smith EDA. He said he knows of people who are taking weeks to run OPC jobs. "It appears we're starting to see the death of general-purpose servers," he said. But one problem with the Cell, he said, is the difficulty of programming its multicore architecture.
"The computing that needs to be done to design electronic systems requires high-performance computing," noted Tom Kingsley, product-marketing manager for resolution enhancement technology at Mentor Graphics.
Specialized hardware for OPC isn't new. It's also offered by Brion Technologies Inc. with Tachyon, a hardware accelerator for lithography modeling and database handling. "It's exciting to see the industry moving toward the image-based, hardware-accelerated approach to computational lithography," said Jim Wiley, senior technical director at Brion.
From games to chip design
IBM Corp.'s multicore Cell architecture is "uniquely suited" to tackling OPC, Sawicki of Mentor said. Originally aimed at gaming applications, the Cell contains one PowerPC processor and eight "synergistic processing elements." The Cell's strength is rapid image processing. Compared with an Opteron processor, a Cell processor uses fast Fourier transforms (FFTs) to speed OPC simulation, he said.
Mercury Computer is a provider of computing systems and software for data-intensive applications, including image processing, signal processing and visualization. In 2005, Mercury and IBM announced an alliance to develop and commercialize Cell BE processor-based products. Mercury followed up with the release of the Dual Cell-based Blade last year for selected compute-intensive applications. The product is just now beginning to go into production use.
The Dual Cell-Based Blade offers peak performance of 400 gigaflops, features two 3.2-GHz Cell BE processors and includes 512 Mbytes of XDR DRAM memory per Cell BE processor. Mercury has mapped key signal-processing algorithms onto the blade.
Sawicki noted that Mentor and Mercury worked together to port Mentor's OPC modeling infrastructure onto the Cell, take advantage of parallelization and optimize Calibre nmOPC for the Cell's FFT library.
The Cell BE is not, as Mercury's Skalabrin noted, a general-purpose processor, so porting and optimizing an application takes some work.
New approach to OPC
Calibre nmOPC brings about a number of improvements beyond the ability to run on Cell BE-based hardware. One is a move to grid-based, "dense" simulation. As Sawicki explained, OPC until now has generally used "sparse" simulation that doesn't involve all of the points on a grid. But dense simulation, which uses the entire grid and is thus more accurate, actually becomes more efficient with increasing layout density.
A fourth-generation compact resist process model, called CM1, supports dense simulation and is said to be more accurate than its predecessors. In addition, it has advantages in stability, calibration automation and process window accuracy, the company said.
Calibre nmOPC is available now starting at $120,000. The Dual Cell-Based Blade is available separately from Mercury Computer. Mentor has not yet named any customers for Calibre nmOPC.