本文来源:https://home.nzcity.co.nz/nl.aspx?5272433
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
英伟达概述了其所谓的半导体制造AI定义模型,该模型结合了加速计算、基于物理的人工智能、人工智能代理和数字孪生。
该模型专注于使用人工智能工具来缩短芯片制造中的工程分析时间。在芯片制造中,评估工艺变更、材料相互作用和缺陷可能需要数天时间,而在生产线上进行物理实验的成本也很高。
随着对先进芯片的需求不断增加,以及制造工艺日益复杂,半导体制造商正面临着压力。在这种环境下,即使是产出或工艺质量方面的微小误差也可能造成重大的商业后果。
更快的模拟。该方法的一个核心部分是更快的模拟。加速计算使工程师能够在决定在实际生产中尝试哪些变更之前,先测试更多备选方案。
其中一个主要用例是多物理场仿真,它对热传递、电气行为和机械应力等相互作用进行建模。软件提供商COMSOL正在使用Nvidia的cuDSS库,将半导体制造模拟的速度提高多达10倍。
应用材料公司还正与英伟达合作,开展原子级材料的发现和工艺开发工作。英伟达表示,应用材料公司的Ginestra模拟平台与cuDSS集成后,运行速度比仅使用CPU的方法快10倍,而其ACE+多物理场平台使用英伟达PhysicsNeMo进行地形模拟,速度可提升35倍。
一些工作负载在传统处理器上难以以生产规模运行。计算光刻技术决定了如何将电路图案印在硅片上,这是一个例子,因为计算量非常大。
西门子正在将Nvidia cuLitho集成到其Calibre光刻软件中。Nvidia表示,这可使关键模拟工作负载的运行速度提升多达17倍,为工程师在晶圆进入印刷阶段之前验证设计提供了更大的空间。
基于物理学的人工智能。Nvidia还阐述了基于物理学的人工智能模型的优势,这些模型可以预测材料和制造工艺的表现,而无需每次都进行全方位模拟。Nvidia认为,对于即使在加速硬件上仍要求很高的昂贵计算,这种方法的益处最大。
等离子体建模就是一个例子。等离子体用于半导体蚀刻工艺,对其行为进行建模需要能够捕捉电气、化学和热效应之间相互作用的方程。
据英伟达称,Lam Research已经证明,英伟达PhysicsNeMo框架可以在毫秒内预测等离子体的行为,而进行一次完整模拟则需要数天时间。这为工程师提供了一种更快的探索工艺条件的方法,而无需每次都重新运行完整的模型。
AI 代理。英伟达还强调了 AI 代理,将其作为一种自动化传统上由人工处理的一系列工程任务的方法。在半导体生产中,工程师仍然负责做出决策,但许多中间步骤可能需要数小时才能完成。
一个例子是选择将哪种人工智能模型应用于特定任务。Nvidia表示,Advantest正在使用Nvidia NeMo Switchyard AI代理来自动化该过程,并将响应时间从数小时缩短至数分钟。
数字孪生。数字孪生是拟议制造模型的另一个组成部分。这些虚拟环境使团队能够在生产线上实施之前,先测试设备配置、工厂布局和工艺变更。
据Nvidia称,应用材料公司使用Nvidia Omniverse库来构建整个制造工厂的物理上准确的数字孪生。这些模型用于模拟材料的移动、识别瓶颈以及评估布局、吞吐量、成本和过程控制。
其目标是将模拟的范围从单个制造步骤扩展到整个工厂的运营。随着制造工厂变得越来越复杂,以及流程变更失败的成本不断增加,这一点变得越来越重要。
互联工作流程。Nvidia的更广泛论点是,当这些技术在封闭的工程环路中相互关联时,其效用会更大。在该模型中,生产数据为模拟和人工智能模型提供信息,然后用于测试各种备选方案,再由工程师在车间做出决策。
英伟达将此作为缩短整个半导体制造反馈周期的一种方式。据该公司称,其目的是帮助工程师更早地发现问题,评估更多方案,并更快地根据生产洞察采取行动。
推动这一举措的原因在于,芯片制造商及其供应商正在寻求各种方法来应对日益复杂的工艺流程,而不必完全依赖速度较慢、成本较高的物理试验。
本文来源:https://home.nzcity.co.nz/nl.aspx?5272433
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Nvidia has outlined what it calls an AI-defined model for semiconductor fabrication, combining accelerated computing, physics-based AI, AI agents and digital twins.
The model focuses on using AI tools to shorten engineering analysis in chip manufacturing, where assessing process changes, materials interactions and defects can take days, and physical experiments on production lines are costly.
Semiconductor manufacturers are under pressure as demand for advanced chips rises and manufacturing processes grow more complex. In that environment, even small errors in yield or process quality can have significant commercial consequences.
Faster simulation. A central part of the approach is faster simulation. Accelerated computing lets engineers test more alternatives before deciding which changes to try in physical production.
One of the main use cases is multiphysics simulation, which models interactions such as heat transfer, electrical behaviour and mechanical stress. Software provider COMSOL is using Nvidia's cuDSS library to speed up semiconductor manufacturing simulations by up to 10 times.
Applied Materials is also working with Nvidia on atomic-scale materials discovery and process development. Nvidia says Applied's Ginestra simulation platform runs up to 10 times faster than CPU-only approaches when integrated with cuDSS, while its ACE+ multiphysics platform delivers topography simulations up to 35 times faster with Nvidia PhysicsNeMo.
Some workloads have been difficult to run at production scale on conventional processors. Computational lithography, which determines how circuit patterns are printed onto silicon, is one example because the calculations are highly intensive.
Siemens is integrating Nvidia cuLitho into its Calibre lithography software. Nvidia says this delivers up to 17 times acceleration on key simulation workloads, giving engineers more scope to verify designs before a wafer reaches print.
Physics-based AI. Nvidia also made the case for physics-based AI models that predict how materials and manufacturing processes will behave without running a full simulation each time. It argues the benefit is greatest for expensive calculations that remain demanding even on accelerated hardware.
Plasma modelling is one example. Plasma is used in semiconductor etch processes, and modelling its behaviour requires equations that capture interactions between electrical, chemical and thermal effects.
According to Nvidia, Lam Research has shown that the Nvidia PhysicsNeMo framework can predict plasma behaviour in milliseconds, compared with days for a full simulation. That gives engineers a faster way to explore process conditions without rerunning the complete model each time.
AI agents. Nvidia also highlighted AI agents as a way to automate sequences of engineering tasks traditionally handled manually. Engineers still make the decisions in semiconductor production, but many intermediate steps can consume hours.
One example is selecting which AI model to apply to a particular task. Nvidia says Advantest is using Nvidia NeMo Switchyard AI agents to automate that process and cut response times from hours to minutes.
Digital twins. Digital twins form another part of the proposed manufacturing model. These virtual environments let teams test equipment configurations, fab layouts and process changes before carrying them out on the production line.
Applied Materials uses Nvidia Omniverse libraries to build physically accurate digital twins of entire fabrication plants, according to Nvidia. The models are used to simulate material movement, identify bottlenecks and assess layouts, throughput, cost and process control.
The aim is to extend simulation beyond individual manufacturing steps to the operation of the factory as a whole. That is increasingly relevant as fabrication plants become more complex and the cost of unsuccessful process changes rises.
Connected workflow. Nvidia's broader argument is that these technologies become more useful when linked in a closed engineering loop. In that model, production data informs simulations and AI models, which are then used to test alternatives before engineers make decisions on the factory floor.
Nvidia presented this as a way to shorten feedback cycles across semiconductor manufacturing. The aim, it says, is to help engineers identify problems earlier, evaluate more options and act on production insights sooner.
The push comes as chipmakers and their suppliers look for ways to manage greater process complexity without relying solely on slower, more expensive physical trials.
