{"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/rethinking-floating-point-for-deep-learning","title":"Rethinking floating point for deep learning","arxiv_id":"1811.01721","date":"2018-11-01","proceeding":null,"authors":["Jeff Johnson"],"abstract":"Reducing hardware overhead of neural networks for faster or lower power\ninference and training is an active area of research. Uniform quantization\nusing integer multiply-add has been thoroughly investigated, which requires\nlearning many quantization parameters, fine-tuning training or other\nprerequisites. Little effort is made to improve floating point relative to this\nbaseline; it remains energy inefficient, and word size reduction yields drastic\nloss in needed dynamic range. We improve floating point to be more energy\nefficient than equivalent bit width integer hardware on a 28 nm ASIC process\nwhile retaining accuracy in 8 bits with a novel hybrid log multiply/linear add,\nKulisch accumulation and tapered encodings from Gustafson's posit format. With\nno network retraining, and drop-in replacement of all math and float32\nparameters via round-to-nearest-even only, this open-sourced 8-bit log float is\nwithin 0.9% top-1 and 0.2% top-5 accuracy of the original float32 ResNet-50 CNN\nmodel on ImageNet. Unlike int8 quantization, it is still a general purpose\nfloating point arithmetic, interpretable out-of-the-box. Our 8/38-bit log float\nmultiply-add is synthesized and power profiled at 28 nm at 0.96x the power and\n1.12x the area of 8/32-bit integer multiply-add. In 16 bits, our log float\nmultiply-add is 0.59x the power and 0.68x the area of IEEE 754 float16 fused\nmultiply-add, maintaining the same signficand precision and dynamic range,\nproving useful for training ASICs as well.","url_abs":"http://arxiv.org/abs/1811.01721v1","url_pdf":"http://arxiv.org/pdf/1811.01721v1.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":"rethinking-floating-point-for-deep-learning","repo_url":"https://github.com/facebookresearch/deepfloat","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":1,"framework":"pytorch","reach":null}],"tasks":[{"task_slug":"deep-learning","task_name":"Deep Learning"},{"task_slug":"math","task_name":"Math"},{"task_slug":"quantization","task_name":"Quantization"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1811.01721","mcp":{"get_harvested_code_for_paper":{"arxiv_id":"1811.01721"}},"developers":"https://syntology.ai/developers","read_at":"2026-09-24T18:15:14+00:00","read_at_is":"when the build read Syntology's graph, not when any sample ran","claim":"Per-sample execution status on synthesized fixtures; not a correctness claim about the paper. Samples come from repositories linked to the paper, official or community; repo_kind says which.","repos":[{"provenance":"external:paperswithcode_snapshot_2025-07-28","url":"https://github.com/facebookresearch/deepfloat","reach":null}],"summary":{"ran_draft_wrong":1},"by_repo_kind":{"official":{"samples":1,"ran":1,"repositories":1}},"repo_kind_vocabulary":{"official":"The archive marks this repository official for the paper","named_in_paper":"The archive records that the paper mentions this repository; it is not marked official","listed":"In the archive's code links for this paper, not marked official and not recorded as mentioned in the paper","found_in_text":"Syntology found this repository in the paper's own text; whether it is the authors' implementation is not asserted","community":"Not in the archive's code links for this paper; a community repository Syntology harvested"},"n_pointer_only_for_licence":1,"samples":[{"code_sha256_prefix":"374f14b37091da2d","entry":"get_fpga_mods","repo":"facebookresearch/deepfloat","repo_kind":"official","path":"py/run_fpga_resnet.py","file_url":"https://github.com/facebookresearch/deepfloat/blob/HEAD/py/run_fpga_resnet.py","link_basis":"first_harvest_node","language":"python","status":"ran_draft_wrong","verification_level":1,"contract_check":"OUTPUT_MISDECLARED","metamorphic_tier":"deterministic","behaviour_fingerprint":false,"licence":"NOASSERTION","inline_ok":false,"mcp_get_code":{"code_sha256":"374f14b37091da2d"}}]},"arxiv_metadata":null,"syntology_extracted_results":null}