<?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>Optimisation on smcleod.net</title>
    <link>https://smcleod.net/tags/optimisation/</link>
    <description>Recent content in Optimisation on smcleod.net</description>
    <image>
      <title>smcleod.net</title>
      <url>https://smcleod.net/logo-painted-350.svg</url>
      <link>https://smcleod.net/logo-painted-350.svg</link>
    </image>
    <generator>Hugo -- 0.152.2</generator>
    <language>en</language>
    <lastBuildDate>Wed, 02 Sep 2026 09:01:14 +1000</lastBuildDate>
    <atom:link href="https://smcleod.net/tags/optimisation/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>Patching NVIDIA&#39;s driver and vLLM to enable P2P on consumer GPUs</title>
      <link>https://smcleod.net/2026/02/patching-nvidias-driver-and-vllm-to-enable-p2p-on-consumer-gpus/</link>
      <pubDate>Wed, 25 Feb 2026 23:59:00 +1000</pubDate>
      <guid>https://smcleod.net/2026/02/patching-nvidias-driver-and-vllm-to-enable-p2p-on-consumer-gpus/</guid>
      <description>NVIDIA artificially restricts peer-to-peer (P2P) GPU communication to their enterprise cards. Turns out this is a software limitation, not a hardware one. I patched my drivers to remove it, hacked vLLM to take advantage of it, and got a 15-50% throughput improvement running Qwen 3.5 35b on dual RTX 3090s.</description>
    </item>
  </channel>
</rss>
