Machine Learning Hardware Exploration via Parametric Analysis Software

机器学习（ML）低功耗，高性能，移动智能手机数字硬件的进步为改进训练ML模型的执行创造了无与伦比的机会，从而扩大了移动计算的应用空间。但这些成功也推动了数据带宽和数字操作性能要求的快速扩展，从而增加了半导体硬件开发商（如高通公司）的成本，高通公司是参与该应用程序的工业合作伙伴。许多因素限制了训练ML模型的改进执行，包括低功耗预算和对下一代智能手机、平板电脑、无人机、物联网和其他移动设备的数据带宽和数字性能需求的成本限制。然而，很有可能在所有这些不同的应用空间中，机器学习算法——充分了解现有和新兴智能手机技术中可用的计算细微差别——可能还会被发明出来，以充分解决这一资源挑战，从而大大拓宽复杂推理问题对移动设置的适用性。深入了解这种解决方案是高通的主要研究兴趣。鉴于这些资源问题，越来越需要了解ML算法到各种可用移动硬件的详细映射。通过先进的参数分析软件（PAS），可以准确地评估各种机器学习架构在任何硬件平台上的性能，并提供足够有洞察力的指标，为未来的微架构设计选择提供信息。然而，机器学习行业缺乏PAS。本研究项目旨在建立一个框架，使用严格的软件和硬件评估技术来评估各种各样的机器学习微架构硬件。

Advances in machine learning (ML) low-power, high-performance, mobile smartphone digital hardware have created unparalleled opportunities for improving the execution of trained ML models and thus expanding the application space of mobile computing. But these successes have also driven rapid expansion of data bandwidth and digital operational performance requirements thus increasing the costs to semiconductor hardware developers such as Qualcomm, the industrial partner involved in this application. Many factors limit the improved execution of trained ML models including low power budgets and cost restrictions placed on data bandwidth and digital performance demands of next generation smartphone, tablet, drone, IoT and other mobile devices. Yet, it is highly possible that across all of these diverse application spaces, ML algorithms - thoroughly informed of the computational nuances available in existing and emerging smartphone technologies - may yet be invented to adequately address this resourcing challenge and as a result greatly broaden the applicability of sophisticated inference problems to mobile settings. Getting insights into such a solution is of prime research interest for Qualcomm. In light of these resourcing issues there is an increasing need to understand the detailed mapping of ML algorithms onto the diverse mobile hardware available. An effective means of gaining such insights is possibly available via advanced parametric analysis software (PAS) designed to accurately assess the performance of various ML architectures on any hardware platform and provide metrics insightful enough to inform future micro-architectural design choices. However, there is a lack of PAS in the machine learning industry. This research project aims to build a framework for evaluation of a wide variety of ML micro-architectural hardware using rigorous software and hardware evaluation techniques.