<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Need for VRAM</title><link>https://needforvram.com/</link><description>The essentials for people who do not intend to miss the AI wave: what to run, on what, at what cost, with whom, and where the community meets. The name is the old phrase about the sinews of war - not the guns, the thing that decides whether you can fight at all.</description><language>en</language><lastBuildDate>Mon, 19 Oct 2026 14:42:05 +0200</lastBuildDate><atom:link href="https://needforvram.com/index.xml" rel="self" type="application/rss+xml"/><item><title>The architecture ladder: why a card that fits can still be refused</title><link>https://needforvram.com/gpus/architecture-ladder/</link><guid isPermaLink="true">https://needforvram.com/gpus/architecture-ladder/</guid><pubDate>Wed, 07 Oct 2026 00:00:00 +0000</pubDate><description>Compute capability 8.9, AMD gfx targets, Intel generations, Apple GPU families: what each step of GPU hardware added, which engines refuse below a step, and how to read your own card's step in one command.</description></item><item><title>The 48 GB RTX 4090</title><link>https://needforvram.com/gpus/4090-48gb/</link><guid isPermaLink="true">https://needforvram.com/gpus/4090-48gb/</guid><pubDate>Fri, 02 Oct 2026 00:00:00 +0000</pubDate><description>NVIDIA never sold one. Who takes a 24 GB card to 48 GB, what it costs, what it breaks, and how to check a modified card before the money leaves your account.</description></item><item><title>Test a new drive before you trust it</title><link>https://needforvram.com/storage/test-a-new-drive/</link><guid isPermaLink="true">https://needforvram.com/storage/test-a-new-drive/</guid><pubDate>Fri, 02 Oct 2026 00:00:00 +0000</pubDate><description>The exact commands to run on a brand-new SSD or hard drive: SMART baseline, capacity verification, a full-surface write and read, and how to tell a used drive from a new one.</description></item><item><title>How drives get faked</title><link>https://needforvram.com/storage/counterfeit-drives/</link><guid isPermaLink="true">https://needforvram.com/storage/counterfeit-drives/</guid><pubDate>Fri, 02 Oct 2026 00:00:00 +0000</pubDate><description>Fake capacity, copied labels, wiped usage logs, drives off a Chia farm: how storage fraud actually works, the channels it arrives through, and the checks that still catch it.</description></item><item><title>Choosing a drive for models and datasets</title><link>https://needforvram.com/storage/choosing-an-nvme/</link><guid isPermaLink="true">https://needforvram.com/storage/choosing-an-nvme/</guid><pubDate>Fri, 02 Oct 2026 00:00:00 +0000</pubDate><description>TLC versus QLC, DRAM versus HMB, the sustained-write cliff, used datacenter U.2 drives: which numbers about an SSD survive contact with a local AI workload.</description></item><item><title>Which quant should I pick?</title><link>https://needforvram.com/pit-stop/which-quant/</link><guid isPermaLink="true">https://needforvram.com/pit-stop/which-quant/</guid><pubDate>Thu, 01 Oct 2026 00:00:00 +0000</pubDate><description>Q4_K_M, IQ4_XS, Q6_K: what the names mean and where quality starts to drop.</description></item><item><title>Two different GPUs, one model</title><link>https://needforvram.com/tuning/mixed-gpus/</link><guid isPermaLink="true">https://needforvram.com/tuning/mixed-gpus/</guid><pubDate>Thu, 01 Oct 2026 00:00:00 +0000</pubDate><description>How llama.cpp splits a model across mismatched cards, and what it means for speed.</description></item><item><title>Power-limit your card, keep your speed</title><link>https://needforvram.com/tuning/power-limit/</link><guid isPermaLink="true">https://needforvram.com/tuning/power-limit/</guid><pubDate>Thu, 01 Oct 2026 00:00:00 +0000</pubDate><description>Generation speed depends on memory bandwidth far more than on GPU power. A lower power limit often costs very little.</description></item><item><title>Out of memory: what to try, in order</title><link>https://needforvram.com/pit-stop/out-of-memory/</link><guid isPermaLink="true">https://needforvram.com/pit-stop/out-of-memory/</guid><pubDate>Thu, 01 Oct 2026 00:00:00 +0000</pubDate><description>The model does not load, or crashes on a long prompt. Seven fixes, from the cheapest to the most costly.</description></item><item><title>YouTube channels</title><link>https://needforvram.com/channels/</link><guid isPermaLink="true">https://needforvram.com/channels/</guid><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><description>The English-language YouTube channels worth following if you run models on your own hardware: local inference, GPU and VRAM hardware, self-hosting models, and the people who measure them.</description></item><item><title>Will it run?</title><link>https://needforvram.com/will-it-run/</link><guid isPermaLink="true">https://needforvram.com/will-it-run/</guid><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><description>Pick a card and see which engines refuse it and which ones only lose their fast paths. Every floor comes from the engine's own code or documentation, with the line quoted.</description></item><item><title>Will it fit?</title><link>https://needforvram.com/will-it-fit/</link><guid isPermaLink="true">https://needforvram.com/will-it-fit/</guid><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><description>Pick your cards, a model, a quant and a context length. See what fits in VRAM, how much room is left, and the speed ceiling of your setup.</description></item><item><title>Threadripper with a mismatched pair</title><link>https://needforvram.com/garage/needforvram-threadripper/</link><guid isPermaLink="true">https://needforvram.com/garage/needforvram-threadripper/</guid><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate></item><item><title>Submit</title><link>https://needforvram.com/submit/</link><guid isPermaLink="true">https://needforvram.com/submit/</guid><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><description>Know someone, a model, a company or an event that belongs here? Send it over.</description></item><item><title>Runtimes</title><link>https://needforvram.com/runtimes/</link><guid isPermaLink="true">https://needforvram.com/runtimes/</guid><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><description>The inference engines that actually run models on your own hardware, and the gateways in front of them: local runtimes, serving engines, routing layers, typed-decision models, diffusion, speech and the interfaces on top.</description></item><item><title>People</title><link>https://needforvram.com/people/</link><guid isPermaLink="true">https://needforvram.com/people/</guid><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><description>Builders, researchers, quantizers and hardware tinkerers worth following. Links go straight to their profiles, with nothing embedded.</description></item><item><title>Models</title><link>https://needforvram.com/models/</link><guid isPermaLink="true">https://needforvram.com/models/</guid><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><description>Open-weight models worth knowing, sorted by how much VRAM they need at 4-bit. Pick your card size to see what fits.</description></item><item><title>Meme machine</title><link>https://needforvram.com/scenes/</link><guid isPermaLink="true">https://needforvram.com/scenes/</guid><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><description>One button, one GPU render. The machine picks a scene and up to three captions about running models at home and the arguments around AI, and paints them on a photorealistic picture. There is no prompt box because you send nothing at all.</description></item><item><title>Legal notice</title><link>https://needforvram.com/legal/</link><guid isPermaLink="true">https://needforvram.com/legal/</guid><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><description>Publisher, host and personal data information, as required by French law (LCEN) and the GDPR.</description></item><item><title>Leaderboard</title><link>https://needforvram.com/leaderboard/</link><guid isPermaLink="true">https://needforvram.com/leaderboard/</guid><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><description>Measured generation speeds on real home hardware. Same method for every run, so the numbers can be compared.</description></item><item><title>Glossary</title><link>https://needforvram.com/glossary/</link><guid isPermaLink="true">https://needforvram.com/glossary/</guid><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><description>The words that decide whether a model runs: video memory, bandwidth, quantisation, KV cache, MoE, prefill and decode, explained without hand-waving.</description></item><item><title>Events</title><link>https://needforvram.com/events/</link><guid isPermaLink="true">https://needforvram.com/events/</guid><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><description>Conferences, trade shows and Maker Faires coming up. Filter by world region and country, or add the whole list to your own calendar. Every date is written as day, month and year.</description></item><item><title>Consulting</title><link>https://needforvram.com/consulting/</link><guid isPermaLink="true">https://needforvram.com/consulting/</guid><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><description>Sizing a machine for a workload, or paying too much for inference? Write in and get introduced to people who do this for a living.</description></item><item><title>Companies</title><link>https://needforvram.com/companies/</link><guid isPermaLink="true">https://needforvram.com/companies/</guid><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><description>Labs, hardware makers and toolmakers betting on AI that runs on your own machine.</description></item><item><title>Books</title><link>https://needforvram.com/books/</link><guid isPermaLink="true">https://needforvram.com/books/</guid><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><description>Books on running models on your own hardware: inference and serving, quantization, accelerators, transformer internals, ML systems, local deployment. One row per title, each naming the catalogue it came from.</description></item><item><title>An RTX 3090 on a 32-core Threadripper</title><link>https://needforvram.com/garage/rtx-3090-threadripper/</link><guid isPermaLink="true">https://needforvram.com/garage/rtx-3090-threadripper/</guid><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate></item><item><title>Add your rig</title><link>https://needforvram.com/contribute/</link><guid isPermaLink="true">https://needforvram.com/contribute/</guid><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><description>Get your own page in the Garage. You write it, you own it, and you can update it whenever your build changes.</description></item><item><title>About</title><link>https://needforvram.com/about/</link><guid isPermaLink="true">https://needforvram.com/about/</guid><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate><description>What this site is and how it is run.</description></item><item><title>A Radeon AI PRO R9700 next to a 3060</title><link>https://needforvram.com/garage/r9700-3060/</link><guid isPermaLink="true">https://needforvram.com/garage/r9700-3060/</guid><pubDate>Mon, 01 Jan 0001 00:00:00 +0000</pubDate></item></channel></rss>