Deep Learning Hardware: Enabling the next wave of applications and innovation

在过去的几年里，我们已经见证了计算机为我们所做的前所未有的进步。突然间，我们可以可靠地与手机通话，或者查找信息。语言间的在线翻译已经变得非常可靠和有用。有些车辆可以在几乎没有人为监督的情况下自动驾驶，而且不断有关于计算机如何被测试以帮助完成许多困难的任务（如医疗诊断）的新闻报道。所有这些创新背后的一项技术是深度学习，这是一种计算算法，可以让计算机在给定足够的例子的情况下自行学习。深度学习背后的核心技术已经存在了几十年。然而，直到最近，人们才可以访问大量的在线信息，计算硬件性能也达到了允许深度学习首次实际应用的水平。虽然现有的硬件足以实现深度学习的第一次突破，但它不足以维持创新。更糟糕的是，虽然在过去，由于半导体技术的进步，计算硬件的性能可以预见地逐年提高，但由于该技术的基本挑战，这已经不可能了。实现硬件性能提升的一种方法是构建专门的硬件，从而实现深度学习的进一步创新。这项为期五年的研究计划将研究这些专业的硬件设计，使深度学习研究人员能够进一步创新，并使我们更接近真正的智能机器。我们专业硬件背后的核心概念是，它利用了深度学习处理过程中计算的值。结果将是一个可编程的，但专业化的架构，将提供至少三个数量级的性能改进比商品解决方案，并可以定制各种使用场景，从大型安装，如数据中心，移动和嵌入式设备。

Over the past few years we have been witnessing unprecedented advances on what computing can do for us. It is suddenly possible to reliably talk to our phones, or to look up information. Online translation between languages have become very reliable and useful. There are vehicles that can drive themselves with very little human supervision, and there are constant news stories about how computers are being put to test to help with many difficult tasks such as medical diagnosis. The one technology behind all these innovations is Deep Learning, a class of computing algorithms that allow computers to learn on their own given enough examples. The core technology behind Deep Learning has been around for decades. However, it is only recently that access to vast amounts of online information became available and computing hardware performance has reached levels that allowed the first practical applications of Deep Learning. While existing hardware was sufficient for these first breakthroughs in Deep Learning, it is not capable enough to sustain innovation. Even worse, while in the past computing hardware performance was improving predictably over the years due to advances in semiconductor technology, this is no longer possible due to fundamental challenges in this technology. One way to deliver the hardware performance increases that will enable further innovation in Deep Learning is to build specialized hardware. This five year research program will investigate such specialized hardware designs enabling Deep Learning researchers to further innovate and to bring us closer to truly intelligent machines. The core concept behind our specialized hardware is that it takes advantage of the values calculated upon during Deep Learning processing. The result will be a programmable, yet specialized architecture that will deliver at least three orders of performance improvements over commodity solutions and that can be tailored to various use scenarios from large scale installations such as data centers to mobile and embedded devices.