2020/03/24 by Nicolas Weber, Felipe Huici, Weber, Nicolas +1
Computer Science · Decision Sciences · #Distributed #Distributed and Parallel Computing Systems #FOS: Computer and information sciences #Machine Learning (cs.LG) #Parallel #Parallel Computing and Optimization Techniques #Scientific Computing and Data Management #and Cluster Computing (cs.DC)
paper · pdf · doi:10.48550/arxiv.2003.10688
openalex publication_date 2020/03/24 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
Modern high performance computing clusters heavily rely on accelerators to\novercome the limited compute power of CPUs. These supercomputers run various\napplications from different domains such as simulations, numerical applications\nor artificial intelligence (AI). As a result, vendors need to be able to\nefficiently run a wide variety of workloads on their hardware. In the AI domain\nthis is in particular exacerbated by the existence of a number of popular\nframeworks (e.g, PyTorch, TensorFlow, etc.) that have no common code base, and\ncan vary in functionality. The code of these frameworks evolves quickly, making\nit expensive to keep up with all changes and potentially forcing developers to\ngo through constant rounds of upstreaming. In this paper we explore how to\nprovide hardware support in AI frameworks without changing the framework's\nsource code in order to minimize maintenance overhead. We introduce SOL, an AI\nacceleration middleware that provides a hardware abstraction layer that allows\nus to transparently support heterogeneous hardware. As a proof of concept, we\nimplemented SOL for PyTorch with three backends: CPUs, GPUs and vector\nprocessors.\n