|Single-CPU vs. compute grid performance comparison showing average single-run execution times for the cam phaser subsystem statistical analysis.|
Modern automotive subsystems depend on an integrated mix of electrical, mechanical, hydraulic, and software technologies. To help account for this mix of technologies, General Motors Powertrain (GMPT) pioneered a Signal Delivery Subsystem (SDSS) development process to design electronic controls and analyze technology interactions. The SDSS process incorporates robust design principles and moves traditional prototype manufacturing and test functions from hardware to software.
GMPT's SDSS process investigates technology interactions between subsystem components by analyzing two distinct signal paths: from sensors through software and into the plant, and from the plant through software to an actuator. In the feed-forward path, a physical parameter is translated from a sensor measurement to the application software in the electronic controller. In the feedback path, a digital word is translated from the application software to the intended actuator response.
Design teams use the SDSS process to quantify how critical subsystem performance metrics are affected by variation, such as changes in component tolerances, environmental conditions, and aging effects.
Robust design methods require testing of multiple subsystem prototypes, with statistically significant test results often requiring hundreds or even thousands of individual test units. Building and testing this many subsystems within a finite development cycle and budget is not possible. The solution is virtual manufacturing, where the benefits of virtual prototyping and distributed processing are combined to effectively build and test multiple subsystems.
GM uses advanced simulation tools, including the Saber simulator from Synopsys, to create multiple virtual prototypes for testing, where each differs only by a change in parameters within tolerance ranges. This technique parallels a real-world manufacturing line, replacing real systems with accurate simulation models that allow parameter changes between simulation runs. Where hardware prototyping allows only a few systems to be built and tested within a finite development cycle, virtual manufacturing multiplies the capability, making it possible to effectively build and test thousands of individual prototypes.
Distributed processing dramatically improves simulation throughput and makes virtual manufacturing practical. Running several thousand simulations to achieve virtual manufacturing goals stretches far beyond the practical limits of a single CPU. To support this process, GMPT and Synopsys employed a cluster of computers in a distributed processing environment to develop an efficient virtual assembly line. Statistical analysis runs are executed in parallel, scaling simulation performance to the number of CPUs on the compute grid; hundreds of virtual subsystems can be built and tested in a matter of hours. Using various automation techniques, design teams can run simulation, analysis, and report-generation activities 24/7.
A variable valve timing (VVT) subsystem can be used to illustrate the benefits of virtual manufacturing and distributed processing for a mechatronic system. A VVT subsystem improves engine performance by dynamically adjusting the timing for an engine's intake and exhaust valves. One of the most popular ways to create the VVT effect uses a cam phaser subsystem, which interprets information from cam and crankshaft sensors to adjust valve timing.
Cam phaser subsystem performance is affected by various design parameters. Because the subsystem is fairly complex, a statistical analysis of even moderate size executed on a single CPU may take several hours or days to complete, depending on design complexity, the number of statistical simulation runs, the range of analyses, and the configuration of the CPU. With a properly configured distributed processing environment, this same statistical analysis can be performed in a fraction of the single-CPU time.
GMPT used virtual manufacturing techniques to analyze a cam phaser VVT system.
Single-CPU execution time averaged 220.5 s per single run of a statistical analysis while the compute grid averaged 17.8 s. A full statistical analysis of 1000 runs with a single CPU takes more than 61 h; GMPT's compute grid is over 12 times faster, needing less than 5 h. Using the compute grid saves considerable development resources and time while significantly improving production subsystem reliability.
As further validation for the performance benefits of distributed processing in its SDSS development flow, a single engineer at GMPT recently completed 60,000 simulation runs in 2½ weeks. These simulations completed analysis on four subsystems of comparable complexity to the cam phaser subsystem described here. Prior to implementation of virtual manufacturing, a single engineer would complete on average one subsystem level analysis every two months.
William Goodwin and Amar Bhatti of General Motors and Michael Jensen of Synopsys wrote this article for AEI.