Much has been made in the technical press about various approaches for automatically parallelizing general-purpose computing. However, there are occasional outstanding opportunities to create domain-specific solutions that can elegantly and efficiently elevate the performance of mission-critical tasks. Mentor Graphics has found such an opportunity in printed circuit board (PCB) routing with their newly-announced “XtremeAR” tool. They have crafted a system that can accelerate the arduous task of PCB auto-routing using up to 15 networked nodes, turning multi-day turnaround times into overnight iterations.
PCB routing has become a bottleneck in many board-based system designs. Increased levels of integration have led to larger, more complex ICs, such as FPGAs, subsuming more of the functionality on a typical board. For the board layout team, this means chips with more pins and more signal integrity concerns talking to each other through more sophisticated boards, while design decisions, such as pinout specifications, are pushed until (and often seemingly beyond) the last possible minute. The result? We have more complicated board designs with less time in which to do them.
Unlike ASIC layout, PCB routing is still often a manual task. The most ambitious boards, and those with some critical analog signals, typically require both automatic and manual routing in order to get the job done effectively. In 2004, Mentor announced what it calls “Xtreme PCB,” which allows multiple, geographically dispersed designers to simultaneously manually route a single board in real time over a network. This permits teams to work together productively to get a PCB design out the door, even if they work thousands of miles apart. Mentor claims that Xtreme PCB customers have been able to reduce layout cycle time for these manually routed designs by 40-70%.
Image courtesy of Mentor Graphics
Now, Mentor has taken advantage of that same underlying server and database technology to create a multi-node auto-routing system capable of harnessing the power of up to 15 machines across a wide-area network. Mentor says that a network of 15 CPUs plus a server can accelerate auto-routing execution up to 10X. This would represent a very efficient multi-node parallel cluster at 66% efficiency. We expect the system will achieve even higher efficiency with less than the maximum number of nodes. Typically, in a multi-whatever processing system, adding processing elements does not produce a linear reduction in execution time. The overhead of coordination and communication between processors reduces the overall efficiency, and there is a point of diminishing marginal return where it no longer pays to add additional computing resources.
Because Mentor’s new XtremeAR system can operate efficiently even over a wide-area network, multi-national companies can surf time zones, pressing otherwise idle machines into service crunching on routing problems. The router nodes are enabled by floating licenses, so you can make the most efficient use of your organization’s computing equipment without having to over-purchase design tool software.
For large, high-speed digital boards, auto routing with conventional systems can run for days. Optimal results – the most efficiently routed board with the fewest layers, highest completion rate, and best timing performance – usually require many iterations of the routing process. Sometimes, the routing process uncovers sub-optimal pin mapping in devices like FPGAs, which requires a re-spin of the FPGA design and a subsequent re-run of the PCB auto-routing process. This iteration-intensive business benefits enormously from faster turnaround, and shortening the auto-route step from days to hours could conceivably lop weeks off of design cycles, perhaps creating the opportunity for big BOM savings as more design iterations allow teams to find more cost-effective solutions.
Capabilities like XtremeAR’s distributed auto-routing help to cement the chasm between the desktop-based low-cost systems used by many small- and mid-sized development teams for their board design and the high-end, enterprise-class systems sold by larger EDA companies. The cost of ownership may be an order of magnitude more than that for simpler systems, but for larger firms with many active design projects, the potential revenue increase from shorter design cycles and the potential profit improvement from better designs easily more than justifies the investment.