Cadence cuts up DRC for speed

Cadence Design Systems has launched a design-rules checking (DRC) engine that can distribute its workload across multiple servers in a cloud, private farm or mixture of both to speed up physical signoff.

Christen Decoin, product management director at Cadence, said Pegasus reflects a major change in the architecture of a signoff tool away from multithreading towards massive parallelism. He said the change was needed to prevent the final full-chip checks from becoming a bottleneck. “It can run on more than a thousand CPUs if needed,” he said.

To split the workload across processors on a network, Pegasus uses one machine to analyze the structure of the design and the rules to be checked, which can be represented as large computational graphs. It then chops up the workload into smaller sub-graphs that can be allocated to multiple processors in the server farm or cloud. The processing kernels use dataflow techniques to pass data between each other before the result is passed back to the management machine.

Decoin said complications such as coloring rules complicate the partitioning. “In advanced nodes you have connectivity rules that mean you can’t cut the design up easily. If you cut in a non-smart way, you might miss an important rule.”

However, the partitioning for spacing rules is much simpler, he added, as they do not necessarily rely on the connectivity to other parts of the design.

The data is handled as a group of streams. “We expect everybody to do that in the future,” Decoin said.

To reduce overall runtime, the management machine starts passing out jobs as soon as it has interpreted enough data on the design and the rule deck to make a start rather than waiting for the entire database to load.

“We have a unique usage model. The workload can be run on a farm or a third-party cloud,” Decoin added. Large-scale users will use a ‘gigascale license’ to provide them with the ability to distribute work across hundreds of nodes, potentially rented for a matter of hours from a service like AWS or Azure. Smaller-scale license packages for eight or sixteen processors will also be available.

Decoin said Pegasus will handle foundries’ existing rule decks.