<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Multi-Tenancy on Le Site de François</title><link>https://lesitedefrancois.be/en/tags/multi-tenancy/</link><description>Recent content in Multi-Tenancy on Le Site de François</description><generator>Hugo -- gohugo.io</generator><language>en</language><copyright>© 2026 François</copyright><atom:link href="https://lesitedefrancois.be/en/tags/multi-tenancy/index.xml" rel="self" type="application/rss+xml"/><item><title>MIG (Multi-Instance GPU)</title><link>https://lesitedefrancois.be/en/ai/mig/</link><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><guid>https://lesitedefrancois.be/en/ai/mig/</guid><description>&lt;p&gt;&lt;strong&gt;MIG (Multi-Instance GPU)&lt;/strong&gt; is an NVIDIA &lt;strong&gt;GPU&lt;/strong&gt; partitioning mode on datacenter accelerators (e.g. &lt;strong&gt;A100&lt;/strong&gt;, &lt;strong&gt;H100&lt;/strong&gt;) that splits one physical card into up to seven &lt;strong&gt;GPU instances (GIs)&lt;/strong&gt;, each with isolated &lt;strong&gt;streaming multiprocessors&lt;/strong&gt;, memory bandwidth, and &lt;strong&gt;HBM&lt;/strong&gt; capacity. The objective is &lt;strong&gt;higher utilization&lt;/strong&gt; in multi-tenant environments: several smaller models or dev/test workloads share one expensive GPU without time-slicing contention as severe as full-card sharing. Each MIG instance appears to the OS and &lt;strong&gt;CUDA&lt;/strong&gt; as a separate GPU with fixed resources; workloads cannot oversubscribe another instance’s memory. MIG suits &lt;strong&gt;inference&lt;/strong&gt; and modest training more often than massive single-job training that needs the entire GPU and &lt;strong&gt;NVLink&lt;/strong&gt; domain.&lt;/p&gt;</description></item></channel></rss>