<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>GPU on Pengcheng's Blog</title><link>https://pengchengneo.github.io/tags/gpu/</link><description>Recent content in GPU on Pengcheng's Blog</description><generator>Hugo -- gohugo.io</generator><language>zh-CN</language><copyright>© 2026 Pengcheng</copyright><lastBuildDate>Tue, 06 Oct 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://pengchengneo.github.io/tags/gpu/index.xml" rel="self" type="application/rss+xml"/><item><title>Ultra-Scale Playbook 简体中文版：在 GPU 集群上训练大语言模型</title><link>https://pengchengneo.github.io/posts/ultrascale-playbook-cn/</link><pubDate>Tue, 06 Oct 2026 00:00:00 +0000</pubDate><guid>https://pengchengneo.github.io/posts/ultrascale-playbook-cn/</guid><description>Ultra-Scale Playbook 简体中文交互版，包含完整章节、附录和 16 个互动实验，覆盖 DP、TP、SP、CP、PP、EP、GPU 内核与混合精度。</description></item><item><title>AI 芯片架构全景：GPU、TPU、Trainium、WSE 与 LPU</title><link>https://pengchengneo.github.io/posts/ai-chip-architectures-zh/</link><pubDate>Mon, 24 Aug 2026 00:00:00 +0000</pubDate><guid>https://pengchengneo.github.io/posts/ai-chip-architectures-zh/</guid><description>系统比较 NVIDIA GPU、Google TPU、AMD GPU、Cerebras WSE、AWS Trainium 与 Groq LPU 的设计理念、计算与内存架构、扩展方式和软件栈。</description></item><item><title>(手写笔记) Reiner Pope 两期黑板课：从 Token 经济学到 AI 芯片底层</title><link>https://pengchengneo.github.io/posts/reiner-pope-blackboard-lectures/</link><pubDate>Mon, 25 May 2026 00:00:00 +0000</pubDate><guid>https://pengchengneo.github.io/posts/reiner-pope-blackboard-lectures/</guid><description>学习 Dwarkesh Podcast 上 Reiner Pope 两期黑板课：第一期用 Roofline 和 batch size 解释 token 成本、MoE 机架布局、KV cache 与长上下文定价；第二期从逻辑门、MAC、mux 和 systolic array 讲到 GPU/TPU/FPGA/CPU 的架构哲学。</description></item><item><title>设计 AI 芯片的硬件与软件</title><link>https://pengchengneo.github.io/posts/designing-ai-chip-cn/</link><pubDate>Fri, 22 May 2026 00:00:00 +0000</pubDate><guid>https://pengchengneo.github.io/posts/designing-ai-chip-cn/</guid><description>Bjarke Hammersholt Roune（前 Google TPUv3 软件技术负责人）关于如何设计 AI 芯片的完整思考，涵盖脉动阵列、数值格式、结构化稀疏、并行化、网络拓扑、软件流水线等核心主题。</description></item><item><title>How to Scale Your Model：大模型扩展完整指南</title><link>https://pengchengneo.github.io/posts/scaling-book-cn/</link><pubDate>Wed, 29 Apr 2026 00:00:00 +0000</pubDate><guid>https://pengchengneo.github.io/posts/scaling-book-cn/</guid><description>《How to Scale Your Model》完整中文翻译，覆盖 Roofline 分析、TPU/GPU 架构、分片矩阵乘法、Transformer 并行策略、LLaMA 训练与推理、性能分析等全部内容</description></item><item><title>Pallas：JAX 内核语言完整指南</title><link>https://pengchengneo.github.io/posts/pallas-jax-kernel-language-cn/</link><pubDate>Sun, 19 Apr 2026 00:00:00 +0000</pubDate><guid>https://pengchengneo.github.io/posts/pallas-jax-kernel-language-cn/</guid><description>JAX Pallas 文档完整中文版，覆盖快速入门、Grid/BlockSpec、软件流水线、TPU 内核编写（矩阵乘法、稀疏计算、分布式、SparseCore）、Mosaic GPU 内核编写（Blackwell 矩阵乘法、集合通信）、设计笔记等全部内容</description></item><item><title>AI 集群互联拓扑——从 TPU 3D Torus + OCS 到 NVLink 全光互联的未来</title><link>https://pengchengneo.github.io/posts/tpu-gpu-interconnect-topology/</link><pubDate>Wed, 15 Apr 2026 00:00:00 +0000</pubDate><guid>https://pengchengneo.github.io/posts/tpu-gpu-interconnect-topology/</guid><description>深入拆解 TPU 的 3D Torus + OCS 光交换架构如何将 9,216 颗芯片连成一张无交换机网络，对比 NVIDIA 从 NVLink8 到 NVLink576 的交换机扩展路线，以及 Feynman 架构拥抱 OCS 的未来展望</description></item><item><title>GPU vs TPU 全家桶横评：从 A100 到 B300，从 TPU v2 到 v7x</title><link>https://pengchengneo.github.io/posts/nvidia-gpu-comparison/</link><pubDate>Tue, 14 Apr 2026 00:00:00 +0000</pubDate><guid>https://pengchengneo.github.io/posts/nvidia-gpu-comparison/</guid><description>对比 NVIDIA A100&lt;del>B300 七款 GPU 与 Google TPU v2&lt;/del>v7x 七代 TPU 的算力、HBM 显存与带宽、互联带宽，梳理两大阵营的演进脉络</description></item></channel></rss>