arXivDaily arXiv每日学术速递 周一至周五更新
北美退伍军人节(Veterans Day),12 日无更新
arXiv 1902.08018cs.DCcs.SYeess.SY
收藏

GPU加速实时控制回路

GPU Acceleration of Real-Time Control Loops

ASML(阿斯麦)

Mohamed A. Bamakhrama, Alejandro Arrizabalaga, Frank Overman, Jean-Paul Smeets, Kornel van der Sommen, Remko van der Vossen, John Wagensveld

AI总结:

本文提出了一种基于GPU的实时控制系统,用于补偿晶圆变形,以减少制造良率损失,通过结合压缩和混合精度算术等技术,利用GPU加速高计算需求的模型。

摘要:

极紫外(EUV)光刻被认为是提高未来十年晶体管密度的关键因素。在EUV光刻中,13.5 nm的EUV光通过掩模照射到硅晶圆上,每片晶圆大约进行100次曝光,每小时可处理超过100片晶圆。在此过程中,部分光能会转化为热量,导致晶圆变形,从而增加对准误差,最终降低制造良率。为缓解这一问题,我们提出了一种坚固的实时控制系统,利用晶圆热前馈模型来补偿晶圆变形。该模型计算预期的晶圆变形,并通过调整光投影和/或晶圆移动来补偿。然而,该模型的计算需求很高,因此需要在专用硬件上执行以快速完成计算。为此,我们部署了图形处理单元(GPU)来加速计算。为了使计算符合所需的时间预算,我们结合了压缩和混合精度算术等多种技术,以及最近的GPU进展,构建了一个基于GPU的实时控制系统。使用NVIDIA P100 GPU的证明概念实现能够提供33 GB/s的解压吞吐量以及每个GPU持续的198 GFLOP/s的混合精度密集矩阵-向量乘法性能。

英文摘要:

Extreme Ultraviolet (EUV) photolithography is seen as the key enabler for increasing transistor density in the next decade. In EUV lithography, 13.5 nm EUV light is illuminated through a reticle, holding a pattern to be printed, onto a silicon wafer. This process is performed about 100 times per wafer, at a rate of over a hundred wafers an hour. During this process, a certain percentage of the light energy is converted into heat in the wafer. In turn, this heat causes the wafer to deform which increases the overlay error, and as a result, reduces the manufacturing yield. To alleviate this, we propose a firm real-time control system that uses a wafer heat feed-forward model to compensate for the wafer deformation. The model calculates the expected wafer deformation, and then, compensates for that by adjusting the light projection and/or the wafer movement. However, the model computational demands are very high. As a result, it needs to be executed on dedicated HW that can perform computations quickly. To this end, we deploy Graphics Processing Units (GPUs) to accelerate the calculations. In order to fit the computations within the required time budgets, we combine in a novel manner multiple techniques, such as compression and mixed-precision arithmetic, with recent advancements in GPUs to build a GPU-based real-time control system. A proof-of-concept implementation using NVIDIA P100 GPUs is able to deliver decompression throughput of 33 GB/s and a sustained 198 GFLOP/s per GPU for mixed-precision dense matrix-vector multiplication.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑