GPU comparison and VRAM planning for local AI.

VRAMGlass is a free web tool for comparing GPUs and planning the hardware needed to run large language models locally. It combines GPU specifications, model memory requirements per quantization level, and asking-price references from second-hand and retail markets in one place, so a user can pick a model, see what it needs, and find a card that fits the budget.

vramglass.com · GPU ranking · Model catalog · Guides · Methodology · About

English · 简体中文 · 繁體中文 · 日本語

VRAMGlass answers four questions that come up every time someone builds or upgrades a machine for local LLM inference:

• Which GPU should I buy for running LLMs locally? VRAMGlass ranks GPUs for local AI by VRAM, memory bandwidth, compute and price per gigabyte of VRAM, across consumer, workstation, data-center and unified-memory devices.
• How much VRAM does a model need? VRAMGlass lists the memory required by each open-weight model at each quantization level, with the KV cache at a given context length added on top, and shows which GPUs the model fits on.
• What does a used GPU actually cost right now? VRAMGlass tracks median asking prices, 24-hour change and 30-day history for used GPUs on eBay and Xianyu, and new GPU offers on Amazon Japan.
• Is more system RAM a substitute for VRAM? VRAMGlass tracks desktop and server RAM prices by specification and explains what offloading to system memory does and does not buy.

Figures below are as of September 18, 2026 and grow over time.

• 45 GPUs and AI devices, including GeForce RTX 30/40/50 series, RTX PRO and Blackwell workstation cards, NVIDIA A100 and V100, AMD Radeon RX and Radeon PRO, AMD Instinct, Intel Arc Pro, Apple Mac Studio and MacBook Pro with unified memory, AMD Ryzen AI Max machines, NVIDIA DGX Spark, and common VRAM-modified cards such as the RTX 4090 48GB and RTX 2080 Ti 22GB.
• 9 open-weight model families and 44 model versions: Qwen, DeepSeek, Llama, Gemma, GLM, Kimi, Mistral, gpt-oss and Hunyuan, with weight sizes at 1-bit through 8-bit quantization levels and FP16.
• 3 price markets, kept separate: eBay used listings in the United States, Xianyu used listings in China, and Amazon.co.jp new offers in Japan. Prices refresh every 6 hours and each device keeps a 30-day price history.
• 8 RAM specification buckets covering DDR4-2666, DDR4-3200, DDR5-5600 and DDR5-6000 desktop and server sticks, tracked as the price of a single stick.
• 13 guides on VRAM versus system RAM, quantization, multi-GPU setups, Mac unified memory versus discrete GPUs, budget GPUs by VRAM tier, and what runs on 16GB of VRAM.

• Prices are asking prices, not completed sales. VRAMGlass records the median, quartiles and sample count of current listings. When fewer than three valid listings remain after cleaning, no price is published for that round.
• Each market is collected and displayed independently. Market selection is separate from display currency; the currency toggle converts at the daily exchange rate for readability and does not merge the markets.
• Listings are cleaned by per-device rules. Mining cards, accessories, bundles, proxy-purchase offers and outlier prices are excluded before the median is computed. A consistency check keeps a sharp move on few samples out of the price history.
• Memory estimates are formula-based and documented. Weight memory is parameters times bytes per parameter at the chosen quantization, plus a KV cache term for the context length and a fixed runtime allowance. Model parameters come from official Hugging Face configurations and quantized weight sizes from actual GGUF file sizes.
• GPU specifications are traceable. Figures come from vendor spec sheets, cross-checked against the TechPowerUp GPU Database and Wikipedia GPU lists.
• Fit and speed figures are planning aids, not benchmarks. Speed estimates follow single-stream llama.cpp behaviour and are calibrated against published runs, usually within about 20 percent.

Every formula, coefficient and cleaning rule is published on the methodology page.

VRAMGlass runs a free remote MCP server, so Grok Bot, Claude, ChatGPT, Claude Code, Cursor and other MCP clients can look up live GPU prices, VRAM requirements and hardware builds while they answer. A VRAMGlass account is required; connect with an API key or sign in with OAuth.

https://mcp.vramglass.com/mcp

Setup steps, client configuration and the tool list are in vramglass/mcp and the developer docs.

The price cards on VRAMGlass are available as a hosted web component. Two lines of HTML render a live card with median price, 24-hour change and 30-day trend:

<script src="https://vramglass.com/widget/v1/price-card.js" async></script>
<vg-price-card gpu="rtx-3090" market="ebay"></vg-price-card>

For React, Vue and other bundled apps there is an npm package:

npm install @vramglass/widgets

The SDK source in vramglass/widgets is MIT licensed. The hosted data and renderer have separate terms: a visible VRAMGlass attribution and a working source link are required. Build and preview a card at vramglass.com/widget.

VRAMGlass is built for local AI developers, hardware enthusiasts and anyone planning a machine to run open-weight models: people deciding between a used RTX 3090 and a new RTX 5070 Ti, checking whether a 70B model fits on two cards, or working out what a Mac Studio with 128GB of unified memory can run. It is an independent project with no vendor sponsorship. Some outbound retail links may be affiliate links, disclosed on the about page.

• Data corrections and questions: hello@vramglass.com
• Widget and SDK issues: vramglass/widgets issues
• MCP server: vramglass/mcp
• Creator on X: @voltwake