Research On DSP Based OCD Simulation Platform Acceleration

Optical critical dimension measurements (OCD) have been widely used in thesemiconductor manufacturing equipment for over10years, used for the detection andcontrol of the etching process, as well as the photoresist trimming. Compared with otherCD measurement method (CD-SEM, AFM, etc.), OCD can measurement very complexstructure or dielectric-filled samples in non-destructive ways, very high measurementaccuracy can be achieved, the cost and volume is much smaller than the othermeasuring equipment.Software is one of the most important components of the OCD measurementplatform, the performance of the software determine the accuracy and efficiency of themeasurement, and the software part consists of the optimizer and simulator. Theoptimizer modeling one model of the sample, then the simulator calculate the opticalresponse, usually the0-order diffraction, and then compares with the real measurementdata. The optimizer modify the model, the simulator calculated again, and the routine isrepeated, until a good agreement with the measured data, then the parameters of themodel is the key parameters of the real structure.Simulator based on the precise theory of diffraction and a certain algorithm forsolving optical response. The performance of the simulator determines the performanceof the entire platform. The performance of the simulator depended on the algorithm.The rigorous coupled-wave analysis (RCWA) is the most commonly used algorithm, itsolves the Maxwell equations. Usually RCWA always calculate large matrix operations,especially eigenvalue solution of which is time-consuming, and restricts the efficiencyof the simulator.After tried GPU and multi-process operation, we know that the bottleneck is thecomputation of eigenvalues in GPU based RCWA acceleration, and this article is basedon attempts using DSP to accelerate simulation platform. At same time, the absorptionlayer method (SAM) avoid the eigenvalue computation, so better parallel computingcharacteristics can achieved, and can also be used to solve the problem of the efficiencyof the simulation platform, this article also tries to accelerate the SAM by using DSP. DSP itself is not used in general-purpose computing, but with the upgrading of the DSPperformance, multiplication performance and memory access performance has beengreatly improved, and the matrix operations’ performance is highly dependence of theboth, which is the reason for our work, also the most important in software design, adetailed introduction and optimization for this is made in the article. What is more, theDSP extensions is very easy,8DSP can be placed on a single PCIE slot; While theextend of CPU need to extend the entire system, such as multi-way servers; And eachPCIE slot can only place1-2GPU cores, and high-end GPU’s means high powerconsumption. This makes it possible to greatly enhance the performance of the platformwith small increase in cost and power consumption.