{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/crossbow-scaling-deep-learning-with-small","title":"CROSSBOW: Scaling Deep Learning with Small Batch Sizes on Multi-GPU Servers","arxiv_id":"1901.02244","date":"2019-01-08","proceeding":null,"authors":["Alexandros Koliousis","Pijika Watcharapichat","Matthias Weidlich","Luo Mai","Paolo Costa","Peter Pietzuch"],"abstract":"Deep learning models are trained on servers with many GPUs, and training must\nscale with the number of GPUs. Systems such as TensorFlow and Caffe2 train\nmodels with parallel synchronous stochastic gradient descent: they process a\nbatch of training data at a time, partitioned across GPUs, and average the\nresulting partial gradients to obtain an updated global model. To fully utilise\nall GPUs, systems must increase the batch size, which hinders statistical\nefficiency. Users tune hyper-parameters such as the learning rate to compensate\nfor this, which is complex and model-specific.\n  We describe CROSSBOW, a new single-server multi-GPU system for training deep\nlearning models that enables users to freely choose their preferred batch size\n- however small - while scaling to multiple GPUs. CROSSBOW uses many parallel\nmodel replicas and avoids reduced statistical efficiency through a new\nsynchronous training method. We introduce SMA, a synchronous variant of model\naveraging in which replicas independently explore the solution space with\ngradient descent, but adjust their search synchronously based on the trajectory\nof a globally-consistent average model. CROSSBOW achieves high hardware\nefficiency with small batch sizes by potentially training multiple model\nreplicas per GPU, automatically tuning the number of replicas to maximise\nthroughput. Our experiments show that CROSSBOW improves the training time of\ndeep learning models on an 8-GPU server by 1.3-4x compared to TensorFlow.","url_abs":"http://arxiv.org/abs/1901.02244v1","url_pdf":"http://arxiv.org/pdf/1901.02244v1.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"crossbow-scaling-deep-learning-with-small","repo_url":"https://github.com/lsds/Crossbow","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"tf","reach":{"status":"ok","spdx":"Apache-2.0"}}],"tasks":[{"task_slug":"deep-learning","task_name":"Deep Learning"},{"task_slug":null,"task_name":"GPU"}],"methods":[{"method_slug":"crossbow","method_name":"Crossbow"}],"datasets_introduced":[],"methods_introduced":[{"slug":"crossbow","name":"Crossbow","full_name":"Crossbow"}],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1901.02244","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}