{"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/sharp-models-on-dull-hardware-fast-and","title":"Sharp Models on Dull Hardware: Fast and Accurate Neural Machine Translation Decoding on the CPU","arxiv_id":"1705.01991","date":"2017-05-04","proceeding":"EMNLP 2017 9","authors":["Jacob Devlin"],"abstract":"Attentional sequence-to-sequence models have become the new standard for\nmachine translation, but one challenge of such models is a significant increase\nin training and decoding cost compared to phrase-based systems. Here, we focus\non efficient decoding, with a goal of achieving accuracy close the\nstate-of-the-art in neural machine translation (NMT), while achieving CPU\ndecoding speed/throughput close to that of a phrasal decoder.\n  We approach this problem from two angles: First, we describe several\ntechniques for speeding up an NMT beam search decoder, which obtain a 4.4x\nspeedup over a very efficient baseline decoder without changing the decoder\noutput. Second, we propose a simple but powerful network architecture which\nuses an RNN (GRU/LSTM) layer at bottom, followed by a series of stacked\nfully-connected layers applied at every timestep. This architecture achieves\nsimilar accuracy to a deep recurrent model, at a small fraction of the training\nand decoding cost. By combining these techniques, our best system achieves a\nvery competitive accuracy of 38.3 BLEU on WMT English-French NewsTest2014,\nwhile decoding at 100 words/sec on single-threaded CPU. We believe this is the\nbest published accuracy/speed trade-off of an NMT system.","url_abs":"http://arxiv.org/abs/1705.01991v1","url_pdf":"http://arxiv.org/pdf/1705.01991v1.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":"sharp-models-on-dull-hardware-fast-and","repo_url":"https://github.com/kpu/intgemm","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"none","reach":{"status":"unanswered"}}],"tasks":[{"task_slug":null,"task_name":"CPU"},{"task_slug":"decoder","task_name":"Decoder"},{"task_slug":"machine-translation","task_name":"Machine Translation"},{"task_slug":"nmt","task_name":"NMT"},{"task_slug":"translation","task_name":"Translation"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1705.01991","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}