About 65 years have passed since the first electronic computers became available. Three earlier generations have passed away, and the fourth, dominated by multiple cores of general purpose microprocessors, is in its waning days. The combination of the slowdown in Moore's Law and the end of Dennard scaling is bringing the 40-year reign of the general purpose microprocessor to an end. Why are these changes in underlying technology proving insurmountable? How can we design computers to deliver improved performance, especially for computationally demanding tasks like machine learning? 自从第一台电子计算机问世以来,大约 65 年过去了。前三代计算机已经成为过去,而以多核通用微处理器为主的第 四代计算机正处于衰败期。摩尔定律的放缓和丹尼德缩放的终结,使通用微处理器 40 年的统治走向终结。为什么底 层技术的这些变化无法克服?我们如何设计计算机来提高性能,特别是对于像机器学习这样计算要求很高的任务?