• 题目：MAGMA: An Optimization Framework for Mapping Multiple DNNs on Multiple Accelerator Cores
• 时间：2022
• 会议：HPCA
• 研究机构：GIT

本篇论文的主要贡献：
We propose an optimization framework called Multiworkload Multi-accelerator Mapping Explorer (M3E).

1. we develop an efficient encoding scheme to encode the search space of the mapping;
2. we develop several modules to enable the found mapping to effectively orchestrate the data movement across sub-accelerator cores;
3. we enable several commonly used black-box optimization algorithms and two reinforcement learning methods to be leveraged as the underlying optimization methods.

1 Related work

• PREMA develops a mapper for multi-tenant language tasks, however targeting single-core accelerator.
• AI-MT successfully designs a mapper for homogeneous multi-core accelerators and shows performance improvement over vision and language tasks.
• Herald targets heterogeneous multi-core accelerators and systematically analyzes the benefit of heterogeneity in dataflows across the accelerator cores for AR/VR workloads (vision tasks).

先比于其他工作，本篇论文的优势在于：

• Optimization-based mapper to solve the mapping problem, while prior arts focus on manually designing a mapper.
• Both homogeneous and heterogeneous DNN accelerator platforms.
• Diverse spectrum of models across vision, language and recommendation, which exhibit different bandwidth requirements

2 Characteristics of DNN Models

本篇论文将DNN模型分解为三类：视觉任务、推荐任务和自然语言处理任务：

• Vision：最熟悉，包含了卷积、MLP和FC
• Recommendation：MLP、attention、embedding lookup layer
• Language：embedding lookup、MLP、RNN、attention layer

所以实验中包括的网络有

• Vision：ResNet、Squeezenet、Mobilenetv2、VGG16、Inception、Mnasnet、Shufflenet
• Language：Electra、Transformer-xl、BERT、Ctrl、Reformer、Flaubert、Bart
• Recommendation：Wide & deep learning for recommender systems、Deeprecsys: A system for optimizing end-to-end at-scale neural recommendation inference、Deep learning recommendation model for personalization and recommendation systems、Deep interest evolution network for click-through rate prediction、Deep interest network for click-through rate prediction