Accelerating AI Applications using Analog In-Memory Computing: Challenges and Opportunities
Accelerating AI Applications using Analog In-Memory Computing: Challenges and Opportunities
复制标题
DOI：
10.1145/3453688.3461746
复制
发表时间：
2021-06
期刊：
影响因子：
--
通讯作者：
Shravya Channamadhavuni;Sven Thijssen;Sumit Kumar Jha;Rickard Ewetz
中科院分区：
文献类型：
--
作者：
Shravya Channamadhavuni;Sven Thijssen;Sumit Kumar Jha;Rickard Ewetz
线性转换是许多人工智能（AI）应用中的主导计算。电阻横梁阵列的自然乘积和累积特征有望为电阻点产品发动机（DPE）提供前所未有的处理能力，该功能可以使用模拟内存中计算加速近似矩阵矢量乘法。不幸的是，加速AI应用程序的功能正确性可能会因各种错误来源而损害。在本文中，我们将概述最紧迫的鲁棒性挑战，最新解决方案的局限性以及未来的研究机会。
Linear transformations are the dominating computation within many artificial intelligence (AI) applications. The natural multiply and accumulate feature of resistive crossbar arrays promise unprecedented processing capabilities to resistive dot-product engines (DPEs), which can accelerate approximate matrix-vector multiplication using analog in-memory computing. Unfortunately, the functional correctness of the accelerated AI applications may be compromised by various sources of errors. In this paper, we will outline the most pressing robustness challenges, the limitations of state-of-the-art solutions, and future opportunities for research.