SBIR Phase I: Addressing the memory bottleneck in deep neural networks in cloud platforms

这个小企业创新研究（SBIR）第一阶段项目的更广泛的影响/商业潜力将包括为部署在云计算上的机器学习应用程序定义超快速和节能加速器的方式。云计算和机器学习的融合正在塑造我们的日常生活体验。运行在云端并利用机器学习算法的应用程序的例子包括数据挖掘、自然语言处理和模式识别。这三者共同代表了认知计算，由于面向开发人员的api数量庞大且不断增长，访问云的计算能力和开发新应用程序变得越来越容易。这种新的计算潜力被企业用来连接数据，寻找对商业有价值的模式，或改善网络安全。这个小企业创新研究（SBIR）第一阶段项目将定义一种新型硬件加速器，能够将认知计算速度提高几个数量级，同时与最先进的处理器相比降低能耗。该技术快速、节能，但易出现精度低、温度变化敏感性低的问题。在第一阶段，公司将在系统级定义硬件加速器，优化设计以实现超高速和足够的精度，以执行所需的认知计算。同时，通过改进设计，将温度变化和噪声的影响降到最低。最后，将评估新设计的能耗，并与最先进的竞争建筑的整体性能进行比较。

The broader impact/commercial potential of this Small Business Innovation Research (SBIR) Phase I project will consist in defining the way toward an ultra-fast and energy efficient accelerator for Machine Learning applications deployed on cloud computing. The merging of cloud computing and Machine Learning is shaping our everyday life experience. Examples of applications running on the cloud and exploiting Machine Learning algorithms include data mining, natural language processing and pattern recognition. These three together represent cognitive computing and, due to a vast and growing number of APIs for developers, it is becoming easier to access the computational power of the cloud and develop new applications. This new computation potential is used by businesses to connect data and find patterns valuable for commerce or to improve cybersecurity.This Small Business Innovation Research (SBIR) Phase I project will define a new kind of hardware accelerator, able to speed up cognitive computation by orders of magnitude while reducing energy consumption compared with state-of-the-art processors. The proposed technology is fast and energy efficient, but can be prone to low precision and temperature variation sensitivity. During Phase I, the company will define the hardware accelerator at the system level, optimizing the design for ultra-high speed and sufficient precision to carry out the cognitive computation required. At the same time, the effect of temperature variation and noise will be minimized through improved design. Finally, the energy consumption of the new designs will be estimated and compared with the overall performance of state-of-the-art competitive architectures.