文章目录
- 题目：GAMMA: Automating the HW Mapping of DNN Models on Accelerators via Genetic Algorithm
- 时间：2022
- 会议：ICCAD
- 研究机构：GIT
本篇论文的主要贡献： 针对DNN mapping定义了完整映射空间，通过遗传算法实现了DNN的自动mapping
DNN加速器的架构可以分成两部分，硬件资源以及映射策略，其中本文的硬件资源主要是面向NoC互连的PE阵列，而硬件映射主要包含了tiling strategy、computation order、parallelization strategy
The architecture of DNN accelerators is determined by two key components: HW resources and HW mapping strategy
The design of computation order and parallelization strategy is also known as dataflow
- 题目：Neu-NoC A High-efficient Interconnection Network for Accelerated Neuromorphic Systems
- 时间：2018
- 会议：ASPDAC
- 研究机构：杜克大学陈怡然
本篇论文的主要贡献：
- hybrid ring-mesh NoC architecture
- neural network aware mapping algorithm
- multicast transmission scheme
- new type of flit to leverage the sparse feature of feature map
hybrid ring-mesh NoC architecture 下图展示了提出的NoC架构，local采用环结构串起8个PE，global此阿勇meshi结构串起不同的环
Sparsity-aware Traffic Reduction： 类似游程编码，当好几个flit都是all-zero flit时，可以用一个flit替换，其中这个flit记录了一共连续有多少个零
Such a scheme allow only one flit indicate multiple sequential flits of 0‘s in the original flit designs