<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>TensorRT on 边缘视觉工程实录</title><link>https://silicegate.cn/tags/tensorrt/</link><description>Recent content in TensorRT on 边缘视觉工程实录</description><generator>Hugo</generator><language>zh-CN</language><lastBuildDate>Tue, 17 Nov 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://silicegate.cn/tags/tensorrt/index.xml" rel="self" type="application/rss+xml"/><item><title>从 43 到 165 FPS：钱都花在哪</title><link>https://silicegate.cn/posts/pipeline-speedup/</link><pubDate>Tue, 17 Nov 2026 00:00:00 +0000</pubDate><guid>https://silicegate.cn/posts/pipeline-speedup/</guid><description>同一块板同一段视频，43.5→143.9→165.3 FPS。拆账：换引擎只值 ×1.15，大头是零拷贝取流和双线程解耦。附单帧账单和三个坑。</description></item><item><title>框是对的，骨架是错的：一个金标测不出的 bug</title><link>https://silicegate.cn/posts/pose-nms-bug/</link><pubDate>Tue, 17 Nov 2026 00:00:00 +0000</pubDate><guid>https://silicegate.cn/posts/pose-nms-bug/</guid><description>姿态模型 91.9 FPS 达标，但压测前埋着一个 bug：NMS 下标用错，框对骨架错，三级金标全过也测不出。人眼看一眼才抓到。</description></item><item><title>十引擎同板对垒：y26n 双冠的赢法是砍 NMS</title><link>https://silicegate.cn/posts/model-shootout/</link><pubDate>Tue, 17 Nov 2026 00:00:00 +0000</pubDate><guid>https://silicegate.cn/posts/model-shootout/</guid><description>五模型十引擎同一条流水线：y26n INT8 165.7 FPS 双冠，但推理段并不最快——赢在把 NMS 从 2.3ms 砍到 0.45ms。附三反直觉和选型卡。</description></item><item><title>先摸个底：这块 Orin Nano 到底能跑多快</title><link>https://silicegate.cn/posts/board-baseline/</link><pubDate>Tue, 17 Nov 2026 00:00:00 +0000</pubDate><guid>https://silicegate.cn/posts/board-baseline/</guid><description>新板到手别急着上项目。五个检测模型、十个 TensorRT 引擎、同一把尺：INT8 后 85~166 FPS 一张表，附功耗温度和值不值的判断。</description></item></channel></rss>