<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[SolarTempest的博客]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/SolarTempest</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; SolarTempest]]></copyright><item><title><![CDATA[海光 DCU 深度技术报告 · 下篇：系统级优化工程]]></title><link>https://blog.youkuaiyun.com/SolarTempest/article/details/162201141</link><guid>https://blog.youkuaiyun.com/SolarTempest/article/details/162201141</guid><author>SolarTempest</author><pubDate>Tue, 23 Jun 2026 14:00:32 +0800</pubDate><description><![CDATA[以"瓶颈转移"为线索，复盘从MATLAB串行17.4小时到HIP/GPU 4.75秒的13,214倍加速历程。重点展开DCU侧的Kernel设计、H2D/Kernel/D2H管线、rocprof内存瓶颈诊断（MemUnitBusy 99.7%）、Occupancy权衡、hipHostRegister的IOMMU踩坑与三缓存管线分析。最终拆解V10.24时间分布：57% Lustre + 19% HBM + 11% PCIe = 87%硬件极限，讨论如何判断优化走到尽头。]]></description><category></category></item><item><title><![CDATA[海光 DCU 深度技术报告 · 中篇：HIP 编程实战与性能分析]]></title><link>https://blog.youkuaiyun.com/SolarTempest/article/details/162200812</link><guid>https://blog.youkuaiyun.com/SolarTempest/article/details/162200812</guid><author>SolarTempest</author><pubDate>Mon, 22 Jun 2026 13:46:16 +0800</pubDate><description><![CDATA[以完整HIP程序为起点，展开SST气候态计算kernel的设计细节——2D Grid划分、Top-34有序数组插入排序、内存合并访问优化、float→short量化。介绍rocprof"三句话诊断法"，讨论Occupancy与内存瓶颈的权衡
  。复盘三个实战Debug案例：P90公式偏差、HDF5线程不安全、Fortran/C维度陷阱。]]></description><category></category></item><item><title><![CDATA[海光 DCU 深度技术报告 · 上篇：硬件架构与软件生态]]></title><link>https://blog.youkuaiyun.com/SolarTempest/article/details/162200617</link><guid>https://blog.youkuaiyun.com/SolarTempest/article/details/162200617</guid><author>SolarTempest</author><pubDate>Mon, 22 Jun 2026 13:43:20 +0800</pubDate><description><![CDATA[从DCU与GPU的本质区别讲起，逐层拆解Z100L的CU/SIMD/Wavefront并行层次、寄存器文件与Occupancy的关系、五级缓存体系及gfx906 ISA特性。软件侧详解DTK四层栈结构、版本兼容性踩坑经验、与CUDA生态对照，附主机处理器和Lustre文件系统概览。]]></description><category></category></item></channel></rss>