A new era for AI HPC and IC technologies in the transition to an intelligent digital world

Abstract

A New era for AI HPC and IC technologies in the transition to an intelligent digital world The massively parallel processing nature of GPU and AI deep learning accelerator (DLA) architectures has enabled the scaling up of computing power, to handle the massive data and large DNN models. Today’s state of the art AI chip in the market, GA100, has a GPU die size close to the reticle limit with 54 billion transistors; it consists of 288 billion transistors in a 2.5D Interposer integrating GA100 GPU/DLA & HBM3, with up to 2.5 Peta Flops of AI inferencing computing power, in a single package that rivals the world’s top supercomputers from 10 years ago. IC industry has been propelled by exponential growth following Moore’s Law since 1960s to 2000. This IC technology scaling is driven by key factors such as performance, power, perfection (yield/reliability), area, cost, and time to market (PPPACT). Since then, Moore’s Law has slowed down significantly as the result of four main challenges: Power Wall since 2000, Performance Scaling Wall since 2010, Area Scaling Wall since 2010s, and the Cost and Time-to-market Wall since late 2010s. On the other hand, since 2012 there has been big progress in three key areas: big data, AI algorithms, and the advancement of GPU and AI accelerators. These have facilitated the rapid transition into an intelligent digital world. Today’s GPUs/DLAs have enabled AI for trillions of devices, with applications ranging across scientific discoveries, medical and drug discovery, robotics, autonomous driving, smart city, IOT, and industrial applications including in the IC and lithography fields etc. In this talk, we will also show examples of the advancements in AI computing and GPU rendering/ray tracing, as well as their applications in various fields. The rapid transition into an intelligent digital world has also demanded a tremendous increase in computing power. For example, the AI model complexity has increased 30000 times in past five years and is currently doubling every two months. To meet the demand for the exponential growth of computing power beyond Moore’s law, full stack innovations are required, including algorithms, system and chip architecture, and IC technologies and materials. Following the Huang’s Law, GPU AI computing power has increased 2x every year since the beginning of the AI Era – an improvement enabled by full stack innovations. In addition, the increase in transistor counts per chip from scaling, and the lower DPPM reliability requirement by autonomous driving applications, has demanded defect-per-chip to reduce by 2x every two years. Today’s advanced chips need to be below one defect per trillion. We will discuss the challenges and opportunities for advanced IC technologies (in particular lithography) and design improvements to achieve continued scaling. The progress in GPU and AI computing has also provided additional tools to solve the problems in designing and productizing future giga chips. AI applications have been used extensively in various fields: IC design, OPC and mask making, IC processing and control, quality & productivity improvements, defect inspection, yield and functionality analysis etc. CNN and GAN are also used to ensure the patterns processed on wafer matches closely to design intentions. We are at the era of an intelligent digital world empowered by AI. The advancements of GPU AI computing improvements following Huang’s Law has fueled the exponential growth of AI applications. A full stack innovation aided by AI is also essential to help the IC and AI/HPC compute industry meet the demand for increased computing power and improved perfection.