LionHeart, a novel framework developed by researchers from IBM Research Europe, Embedded Systems Laboratory (ESL) at EPFL, and HEIG-VD, is designed to execute Deep Learning (DL) inference workloads using heterogeneous accelerators. The framework addresses the challenges of arithmetic precision and stochasticity in analog domain computations. LionHeart operates by obtaining hybrid analog-digital mappings, showcasing high accuracy and potential for speedup across different Deep Neural Networks (DNNs) and datasets. It optimizes DL workloads for edge scenarios and leverages Analog In-Memory Computing (AIMC) to execute Matrix-Vector Multiplications (MVMs) in constant time complexity.
Introduction to LionHeart: A Mapping Framework for Heterogeneous Systems
LionHeart is a novel framework developed by a team of researchers from IBM Research Europe, Embedded Systems Laboratory (ESL) at Ecole Polytechnique Federale de Lausanne (EPFL), and HEIG-VD. The framework is designed to execute Deep Learning (DL) inference workloads using heterogeneous accelerators. It aims to address the challenges of arithmetic precision and stochasticity in analog domain computations, which can degrade application accuracy over time.
The Functionality of LionHeart
LionHeart operates by obtaining hybrid analog-digital mappings. The mappings derived by LionHeart showcase high accuracy and potential for speedup across different Deep Neural Networks (DNNs) and datasets. The results of full system simulations highlight runtime reductions and energy efficiency gains that exceed 6% with a user-defined accuracy threshold, compared to a fully digital floating point implementation.
The Application of LionHeart in Deep Learning
Deep Learning-based solutions have been applied in various industrial and scientific applications, including computer vision, natural language processing, and speech recognition. However, the significant memory and computing requirements of State-Of-The-Art (SOTA) DL models pose a challenge for deploying DL workloads on edge devices. LionHeart addresses this challenge by optimizing DL workloads for edge scenarios.
The Role of In-Memory Computing in LionHeart
In-Memory Computing (IMC) is a computational approach in which storage and computation are performed at the same physical location, avoiding the well-known memory wall problem. LionHeart leverages Analog In-Memory Computing (AIMC), a type of IMC, to execute Matrix-Vector Multiplications (MVMs), the most dominant operation of many Machine Learning (ML) algorithms, in constant time complexity.
The Challenges and Solutions in Implementing LionHeart
A key drawback of AIMC is its sensitivity to device and circuit-related variations, which can adversely affect accuracy. LionHeart addresses this accuracy challenge by exploring hybrid digital-analog mappings of DL models, effectively navigating the tradeoff between runtime performance and accuracy degradation.
The Contributions of LionHeart
The LionHeart framework presents a novel accuracy-driven training framework able to explore the space of hybrid digital-analog implementations of DNNs. It allocates Fully Connected (FC) and Convolutional (CONV) layers to digital or analog resources to minimize runtime while constraining the adverse effect of analog computations on accuracy.
The Future of LionHeart
The LionHeart framework is hardware-agnostic, meaning it can be applied to AIMC crossbars adopting different device technologies. It also addresses the real-world system accuracy degradation caused by temporal variation of programmed AIMCs crossbars at a user-desired evaluation time. The team behind LionHeart continues to investigate these effects and propose strategies to mitigate them.
The article titled “LionHeart: A Layer-based Mapping Framework for Heterogeneous Systems with Analog In-Memory Computing Tiles” was published on January 17, 2024. The authors of this article are Corey Lammie, Flavio Ponzina, Yuxuan Wang, Joshua P. Klein, Marina Zapater, Irem Boybat, Abu Sebastian, Giovanni Ansaloni, and David Atienza. The article was published on arXiv, a platform managed by Cornell University. The DOI reference for this article is https://doi.org/10.48550/arxiv.2401.09420.