<?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>nvlink on Sparse Notes</title>
    <link>https://sparsenotes.com/tags/nvlink/</link>
    <description>Recent content in nvlink on Sparse Notes</description>
    <image>
      <url>https://sparsenotes.com/images/og-default.png</url>
      <link>https://sparsenotes.com/images/og-default.png</link>
    </image>
    <generator>Hugo -- gohugo.io</generator>
    <lastBuildDate>Sat, 03 Oct 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://sparsenotes.com/tags/nvlink/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Mixture-of-Kittens: MoE Megakernel for NVL72s</title>
      <link>https://sparsenotes.com/posts/2026/10/papers/mixture-of-kittens-moe-megakernel-nvl72/</link>
      <pubDate>Sat, 03 Oct 2026 00:00:00 +0000</pubDate>
      
      <guid>https://sparsenotes.com/posts/2026/10/papers/mixture-of-kittens-moe-megakernel-nvl72/</guid>
      <description>Cursor and Stanford rebuild MoE training for a 72-GPU NVLink domain: pull dispatch, push combine, a tunable overlap granularity, and a device-side ring buffer, all inside one deterministic megakernel. Up to 2.37x over the best public baseline, 1.41x end-to-end on 512 GB300s.</description>
    </item>
    
  </channel>
</rss>
