Optimization Schemes for Large Scale Digital Circuits in Presence of Fabrication Randomness

在大规模数字电路中，制造尺寸的减小导致与制造参数相关的随机性显著增加。这已经开始严重影响制造产量，从而影响半导体行业的盈利能力。在这项研究中，研究人员专注于开发大规模数字电路合成的正式优化方案，同时积极考虑随机诱导的产量损失作为优化标准。作为一个关键的智力优势，具体的数字电路优化问题的实例，可以建模为凸规划正在研究存在的制造随机性。利用对产量损失函数性质的良好数学理解，正在开发定制的优化算法。其中一些算法利用了问题实例的特殊数学结构（在某些特殊情况下屈服损失函数的凸性）。对于这种情况，研究人员研究了形式凸优化方法的修改（如切割平面/内点等），以提高收敛速度。在这种数学性质不存在的情况下，正在开发有效的启发式方法。正在研究的一个关键议程是如何将这种优化方案与统计估计方法相结合（统计估计方法测量特定解决方案实例的产量损失，并且已知非常缓慢）。提高半导体工业的生产力和盈利能力，提高纳米技术的适用性（其中制造随机性是一个主要问题），并通过研究生和本科生培训改善教学基础设施是这项工作的关键广泛影响。

Reduction in fabrication dimensions has resulted in significant increase in the randomness associated with the fabricated parameters of large scale digital circuits. This has begun to severely impact the manufacturing yield and therefore the profitability of the semiconductor industry. In this research the investigators are focusing on developing formal optimization schemes for synthesizing large scale digital circuits while proactively considering randomness induced yield loss as an optimization criteria.As a key intellectual merit, specific digital circuit optimization problem instances that can be modeled as convex programs are being investigated in presence of fabrication randomness. Using a sound mathematical understanding of the nature of the yield loss function, customized optimization algorithms are being developed. Some of these algorithms leverage the special mathematical structure of the problem instances (convexity of the yield loss function in some special cases). For such instances the investigators study modifications of formal convex optimization methods (like cutting plane/interior point etc.) for improving the rates of convergence. In cases where such mathematical properties do not exist, efficient heuristics are being developed. A key agenda under investigation is how should such optimization schemes be integrated with statistical estimation methods (which measure the yield loss for a specific solution instance and are known to be very slow). Improving the productivity and profitability of the semiconductor industry, improving the applicability of nanotechnology (where manufacturing randomness is a major concern) and improving the teaching infrastructure through graduate and undergraduate student training are key broad impacts of this work.