As the adoption and sophistication of artificial-intelligence (AI) technology continues to expand, it’s increasingly clear that AI and its many derivatives (deep learning, machine learning, etc.) will lead to profound socioeconomic changes of a magnitude the world hasn’t seen since the industrial revolution. While it can sometimes be a controversial topic in terms of its ethical vs. unethical use, and potential impact on global socioeconomics, one thing is certain: AI is, even in its infancy, being effectively deployed to vastly improve and better automate a number of tasks ranging from data collection, communications, robotics/factory automation, automotive design—and to even our own small but mighty industry, EDA, in chip design.
简而言之（并原谅混合的隐喻），Genie出来了瓶子，我们需要拥抱它。在2020年，我们肯定可以期望看到AI芯片和系统的设计和部署持续增长。
帮助公司提供更复杂的AI技术，EDA公司等Mentor, a Siemens business，做两件事：
1. Developing tools to help companies design AI-accelerators faster.
2. Leveraging machine-learning algorithms to improve IC design tools so that they can deliver better results for customers faster.
Let’s take a closer look at these.
Tools for Better, Faster AI/ML design
Most AI starts out as a mathematic algorithm. The developer creates the algorithm in a math tool, translates it to C code, and then runs that algorithm on a processor. Inevitably, they’ll find that parts of the algorithm run optimally, while other parts run more slowly.
That’s why there’s a vast rise (and subsequent investment from the venture community as well as from large corporations) in AI/ML accelerator IP and silicon development. Some algorithms can be forced (by tweaking the C code) into running on existing AI ASSP architectures with performance and power compromises, but others (think of mobile devices, for example) require dedicated logic to run optimally.
为eda工具开发给定应用程序尖叫的正确AI架构，使公司能够使用更高级别的抽象。我们开始在Compult HLS（高水平综合）技术中，为开发AI IP加速器为其系统的系统（SOC）设计而言，我们开始看到增加的业务。这使AI架构师能够开发他们的数学代码，将其转换为C或Systemc，并在硬件与软件中实现其算法的前端。然后，它们可以迅速收敛于理想的架构，而不是试图立即降到RT水平。
When they do converge on the ideal architecture and select the hardware configuration that runs most suitably, by far the best way to test and refine this architecture is to use an emulation solution, such as Mentor’s Veloce emulation platform. This enables teams to not only run the hardened chip design on the emulation system and then fine-tune the software emulation, but emulation can also be connected to the end system, with emulation performing overall ML benchmarking.
Many of these AI platforms will increasingly require high-speed connectivity. As such, we believe that most ML SoCs will in fact be mixed-signal SoCs. This will require tools that can more effectively bridge the digital and analog domains, such as our Symphony AMS simulation solution, which is a vendor-agnostic AMS environment.
Because many of the ASSP AI devices that are at the heart of data centers and communications infrastructure require the highest possible performance, I believe we’ll start to see increased commercialization of silicon photonic devices that bring photonics directly to the silicon.
ML-Enhanced EDA工具，以更好，结果更快
在过去的几年中，Mentor的研发工作人员在很多方面都领导了将ML集成到我们自己的EDA工具中的方式。该公司目前有五种工具产品可商购，利用ML帮助提供更好的结果，并更快地提供更快的效果。让我在改善EDA工具的背景下盯着一瞬间并谈论ml。
众所周知，没有数据，ml是没用的。产生的数据越多，可以调用越多的ml以排序和开发与该数据有意义的东西。我们幸运的是，EDA工具本身产生了巨大的数据。过程制造产生大量数据，用户可以产生自己的专有数据和要求（用于培训）。因此，理论上没有数据不足。事实上，在利用ML的EDA时，问题变为：可以有效地利用哪些数据集以获取什么工具功能？
到目前为止，我们有五种工具利用ML增强功能，其中大部分都是物理设计，验证和制造领域：ML对于图书馆表征和仿真，光学接近校正，光刻模拟和CMP建模。这个eda段是一个明显的开始。
例如，对于光学邻近校正，Calibre OPC在物理设计数据库上运行。每个芯片运行都会产生数十亿个可用于分析的数据点。通过将ML技术置于Calibre OPC，我们可以收集这些芯片周围的数据，然后使用该数据，以便为设计更快地产生更好的结果。
At 7 nm for a critical layer, customers are using up to 8,000 CPUs running for 12 to 24 hours to perform a single run. By using machine learning, we have been able to drop that by a factor of three and constrain the increase in time that would be necessary to produce each of the advanced nodes coming in the future(see figure)。
这只是一个示例，但我们通过其他ML增强工具看到类似的结果。我们目前有两次其他项目在作品中查看了新的方式，即AI / ML可以利用一个看似大量的数据，以更快地提供更好的结果。您当然可以预期在2020年，您将使用更多ML增强的EDA工具来产生更多的创新。
Joseph Savicki是IC EDA的执行副总裁Mentor, a Siemens Business。
