U-M shares 1st place prize at ISPD

Dongjin Lee, a graduate student in electrical engineering, and his advisor Prof. Igor Markov shared 1st place with teams from National Taiwan University and National Chiao Tung University in the 2009 Clock Network Synthesis Contest at the International Symposium on Physical Design, held March 29-April 1 in San Diego, CA. The contest was organized by IBM Austin Research Laboratory.

This year’s contest tested clock distribution network synthesis. Like the prior placement and routing contests, a set of benchmarks was released and teams were charged with producing clock distribution networks.

Specifically, the contest focused on the computer-aided design and optimization of integrated circuits for distributing accurate 2GHz clock signals throughout a semiconductor chip, with picosecond precision and subject to power constraints. The contest initially attracted 27 teams (15 from the US and 11 from abroad), though only 9 teams developed working software. IBM researchers evaluated submitted programs on “hidden” benchmarks derived from IBM’s recent system-on-chip designs, and verified results by electrical circuit simulation.

As detailed in the results and contest summary (see especially page 28), the three winning teams were far ahead of other contestants, and each excelled in a different aspect of performance. Dongjin’s program, called Contango, performed most reliably, while the competitors had slightly better results on some benchmarks and significantly worse results on others. Although runtime did not affect contest scores, Contango ran twice as fast as the programs with which it tied for first place.

Prof. Markov advised Donjin on the project, and said of his work:

“Dongjin’s research combines algorithms for combinatorial and continuous optimization, especially computational geometry, with integrated circuit design — it’s a perfect blend of computer science with computer and electrical engineering, and can have a huge impact on upcoming integrated circuits. While traditional algorithms typically optimize a well-defined closed-form objective function, Dongjin found that the performance of a clock network is difficult to approximate by analytical formulas with sufficient accuracy. Therefore, his software first builds a conservative initial circuit, and then improves it by invoking the circuit simulator many times and carefully analyzing the performance numbers returned.”