---
id: 20261003-T0-09
title: "Fast Polynomial Transcendentals for LLMs：GPU瓶颈转移下的内核优化"
title_en: "Fast Polynomial Transcendentals Target Shifting GPU Bottlenecks in LLMs"
url: https://ai.daily.yangsir.net/daily/20261003-T0-09
issue_date: 2026-10-03
publish_date: 2026-10-02T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2610.00049
---

# Fast Polynomial Transcendentals for LLMs：GPU瓶颈转移下的内核优化

不同代 GPU 在矩阵、特殊函数和内存管线上的扩展速度不同，导致内核瓶颈随硬件演进发生转移。FlashAttention-4 在 NVIDIA Blackwell 上暴露了注意力内部的不平衡。研究测试了快速多项式超越函数在 LLM 计算中的应用，针对特殊函数计算成为瓶颈的场景，用多项式近似替代高成本超越函数运算。该方法可用于注意力、归一化等环节，在新型 GPU 上提升 LLM 推理和训练的内核效率。

## English Version

**Fast Polynomial Transcendentals Target Shifting GPU Bottlenecks in LLMs**

GPU generations scale matrix, special-function, and memory pipelines at different rates, so kernel bottlenecks shift as hardware evolves. FlashAttention-4 exposed this imbalance inside attention on NVIDIA Blackwell. The paper tests fast polynomial transcendentals for LLM computation, replacing costly transcendental function evaluations with polynomial approximations when special-function computation becomes the bottleneck. The approach can be applied to attention and normalization, improving kernel efficiency for LLM inference and training on newer GPUs.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2610.00049)

**详情页**：https://ai.daily.yangsir.net/daily/20261003-T0-09

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*