<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/">
  <channel>
    <title>LLM-추론-비용 on gnosyslambda&#39;s log</title>
    <link>https://gnosyslambda.github.io/tags/llm-%EC%B6%94%EB%A1%A0-%EB%B9%84%EC%9A%A9/</link>
    <description>Recent content in LLM-추론-비용 on gnosyslambda&#39;s log</description>
    <image>
      <title>gnosyslambda&#39;s log</title>
      <url>https://gnosyslambda.github.io/images/og-image.png</url>
      <link>https://gnosyslambda.github.io/images/og-image.png</link>
    </image>
    <generator>Hugo</generator>
    <language>ko</language>
    <lastBuildDate>Sat, 11 Jul 2026 22:45:30 +0900</lastBuildDate>
    <atom:link href="https://gnosyslambda.github.io/tags/llm-%EC%B6%94%EB%A1%A0-%EB%B9%84%EC%9A%A9/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>LLM 추론 비용 최적화: KV 캐시와 메모리 병목</title>
      <link>https://gnosyslambda.github.io/posts/2026-07-11-llm-inference-memory-bandwidth/</link>
      <pubDate>Sat, 11 Jul 2026 22:45:30 +0900</pubDate>
      <guid>https://gnosyslambda.github.io/posts/2026-07-11-llm-inference-memory-bandwidth/</guid>
      <description>LLM 추론 비용의 핵심 병목을 FLOPs가 아닌 KV 캐시와 GPU 메모리 대역폭 관점에서 분석합니다. 프리필·디코드 분리와 운영 판단 기준을 정리합니다.</description>
    </item>
  </channel>
</rss>
