当前位置: X-MOL 学术arXiv.cs.AR › 论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
DANCE: Differentiable Accelerator/Network Co-Exploration
arXiv - CS - Hardware Architecture Pub Date : 2020-09-14 , DOI: arxiv-2009.06237
Kanghyun Choi, Deokki Hong, Hojae Yoon, Joonsang Yu, Youngsok Kim, Jinho Lee

To cope with the ever-increasing computational demand of the DNN execution, recent neural architecture search (NAS) algorithms consider hardware cost metrics into account, such as GPU latency. To further pursue a fast, efficient execution, DNN-specialized hardware accelerators are being designed for multiple purposes, which far-exceeds the efficiency of the GPUs. However, those hardware-related metrics have been proven to exhibit non-linear relationships with the network architectures. Therefore it became a chicken-and-egg problem to optimize the network against the accelerator, or to optimize the accelerator against the network. In such circumstances, this work presents DANCE, a differentiable approach towards the co-exploration of the hardware accelerator and network architecture design. At the heart of DANCE is a differentiable evaluator network. By modeling the hardware evaluation software with a neural network, the relation between the accelerator architecture and the hardware metrics becomes differentiable, allowing the search to be performed with backpropagation. Compared to the naive existing approaches, our method performs co-exploration in a significantly shorter time, while achieving superior accuracy and hardware cost metrics.

中文翻译:

DANCE:可微加速器/网络协同探索

为了应对 DNN 执行不断增长的计算需求,最近的神经架构搜索 (NAS) 算法考虑了硬件成本指标,例如 GPU 延迟。为了进一步追求快速、高效的执行,DNN 专用硬件加速器被设计用于多种用途,远远超过 GPU 的效率。然而,这些与硬件相关的指标已被证明与网络架构呈现非线性关系。因此,针对加速器优化网络,或针对网络优化加速器,就变成了鸡和蛋的问题。在这种情况下,这项工作提出了 DANCE,这是一种对硬件加速器和网络架构设计进行共同探索的可区分方法。DANCE 的核心是一个可微分的评估者网络。通过使用神经网络对硬件评估软件进行建模,加速器架构和硬件指标之间的关系变得可区分,从而允许通过反向传播执行搜索。与朴素的现有方法相比,我们的方法在更短的时间内执行共同探索,同时实现了卓越的准确性和硬件成本指标。
更新日期:2020-09-15
down
wechat
bug