BAPS: A Fine-Grained Low-Precision Scheme for Softmax in Attention via Block-Aware Precision reScaling
Zisheng Ye, Xiaoyu He, Maoyuan Song, Guoliang Qiu, Chao Liao, Chen Wu, Yonggang Sun, Zhichun Li, Xiaoru Xie, Yuanyong Luo, Hu Liu, Pinyan Lu, Heng Liao · 2026
Read BAPS: A Fine-Grained Low-Precision Scheme for Softmax in Attention via Block-Aware Precision reScaling by Zisheng Ye, Xiaoyu He, Maoyuan Song, Guoliang Qiu, Chao Liao, Chen Wu, Yonggang Sun, Zhichun Li, Xiaoru Xie, Yuanyong Luo, Hu Liu, Pinyan Lu, Heng Liao (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-02071.txt.
Description
Zisheng Ye and colleagues introduce BAPS, a low-precision softmax scheme via block-aware rescaling, addressing the bottleneck. The bottleneck stems from limited bandwidth and exponentiation area cost. Their workflow employs an 8-bit floating-point format and block-aware rescaling, making low-precision softmax feasible. Specifically, the design halves data movement bandwidth by constraining outputs to 8-bit, and reduces exponentiation area by computing in 8-bit precision. Evaluation on language and multi-modal models confirms the method, doubling throughput without increasing chip area.
Catalog record
| CREATOR | Zisheng Ye, Xiaoyu He, Maoyuan Song, Guoliang Qiu, Chao Liao, Chen Wu, Yonggang Sun, Zhichun Li, Xiaoru Xie, Yuanyong Luo, Hu Liu, Pinyan Lu, Heng Liao |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 47317 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602.02071