<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>FP32 on 111qqz's blog</title><link>https://111qqz.com/en/tags/fp32/</link><description>Recent content in FP32 on 111qqz's blog</description><generator>Hugo -- gohugo.io</generator><language>en</language><managingEditor>hust.111qqz@gmail.com (111qqz)</managingEditor><webMaster>hust.111qqz@gmail.com (111qqz)</webMaster><copyright>© 2011-2026 111qqz</copyright><lastBuildDate>Tue, 17 Nov 2026 21:25:00 +0800</lastBuildDate><atom:link href="https://111qqz.com/en/tags/fp32/index.xml" rel="self" type="application/rss+xml"/><item><title>Kahan Summation：用一个补偿变量挽回 FP32 累加丢失的精度</title><link>https://111qqz.com/en/post/%E6%B7%B1%E5%BA%A6%E5%AD%A6%E4%B9%A0/2026-11-17-kahan-summation/</link><pubDate>Tue, 17 Nov 2026 21:25:00 +0800</pubDate><author>hust.111qqz@gmail.com (111qqz)</author><guid>https://111qqz.com/en/post/%E6%B7%B1%E5%BA%A6%E5%AD%A6%E4%B9%A0/2026-11-17-kahan-summation/</guid><description>&lt;p&gt;上一篇&lt;a href="https://111qqz.com/en/post/%E6%B7%B1%E5%BA%A6%E5%AD%A6%E4%B9%A0/2026-11-16-numerical-stability-ml-systems/" &gt;从浮点误差谈到 GPU 计算&lt;/a&gt;时，我提到了 H20 上的精度取舍：量化特征要做横截面累加，而 FP64 吞吐受限，直接用 FP32 又需要留意累加误差。Kahan 补偿求和看起来是个值得考虑的工具，但当时我把它的能力说得太满了。&lt;/p&gt;</description></item></channel></rss>