<?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>GPU on Wiliam Rosa | Lakehouse, Dados &amp; IA</title>
    <link>https://wiliamrosa.github.io/tags/gpu/</link>
    <description>Recent content in GPU on Wiliam Rosa | Lakehouse, Dados &amp; IA</description>
    <image>
      <title>Wiliam Rosa | Lakehouse, Dados &amp; IA</title>
      <url>https://wiliamrosa.github.io/images/wiliam-rosa-blog-og.png</url>
      <link>https://wiliamrosa.github.io/images/wiliam-rosa-blog-og.png</link>
    </image>
    <generator>Hugo</generator>
    <language>pt-br</language>
    <lastBuildDate>Sat, 05 Sep 2026 09:00:00 -0300</lastBuildDate>
    <atom:link href="https://wiliamrosa.github.io/tags/gpu/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>A Databricks construiu um agente que escreve e valida kernel de GPU sozinho</title>
      <link>https://wiliamrosa.github.io/posts/proteus-geracao-kernel-gpu-especializado/</link>
      <pubDate>Sat, 05 Sep 2026 09:00:00 -0300</pubDate>
      <guid>https://wiliamrosa.github.io/posts/proteus-geracao-kernel-gpu-especializado/</guid>
      <description>O Proteus propõe, valida e faz benchmark de kernel de GPU especializado por formato de operação, com um estudo de caso no Gated DeltaNet do Qwen 3.5 122B atingindo de 1,8x a 5,2x de speedup sobre a melhor implementação do vLLM.</description>
    </item>
    <item>
      <title>GPU chegou ao compute serverless do Databricks com o AI Runtime</title>
      <link>https://wiliamrosa.github.io/posts/azure-databricks-ai-runtime-gpu-serverless/</link>
      <pubDate>Fri, 20 Mar 2026 09:00:00 -0300</pubDate>
      <guid>https://wiliamrosa.github.io/posts/azure-databricks-ai-runtime-gpu-serverless/</guid>
      <description>AI Runtime traz suporte a GPU pro serverless em preview público, pra tarefas single-node, a promessa é treinar sem gerenciar cluster, mas ainda com pé atrás no treinamento distribuído.</description>
    </item>
  </channel>
</rss>
