LLM GPU Checker (KO)

Match LLMs to GPUs and plan multi-model AI stacks by VRAM, bandwidth and precision.

Free· Free (open web tool hosted on GitHub Pages).EvaluationCatalog covers open models on Hugging Face (Llama, Qwen, Mistral, Gemma, etc.)

Visit website →

Best for

Engineers and MLOps teams planning self-hosted LLM deployments on their own GPUs who need a quick VRAM/precision sanity check before procurement or a vLLM/Ollama rollout.

Skip if

Teams that only consume hosted APIs (OpenAI, Anthropic, Bedrock) and never touch GPU sizing, or anyone needing audited, guaranteed benchmarks for compliance.

LLM GPU Checker (KO) is a free, browser-based capacity-planning tool that helps engineers work out which open-source LLMs will actually run on a given GPU (or set of GPUs), and vice versa. Built by a Korean developer and hosted on GitHub Pages, it offers a bilingual (Korean / English) interface aimed at practitioners standing up self-hosted inference — the audience running vLLM, llama.cpp, Ollama or plain Transformers on their own hardware rather than paying per-token API fees.

The app has four main workflows. GPU-to-Model lets you pick a card (for example an RTX 4090, A100 40GB or H100) and see which models fit at various quantisation levels. Model-to-GPU works the other way: choose a model such as Llama 3.1 70B or Qwen 2.5 and get the minimum viable hardware. A Service Planner takes a workload description (service type, expected concurrent users, context length) and returns an equipment estimate. Finally, an AI Stack Planner distributes LLMs, embedding models, rerankers and OCR/VLM components across multiple GPUs using pipeline, independent-service or sequential execution modes — useful when you are architecting a full RAG or agent stack rather than a single chatbot.

Calculations key off VRAM footprint and memory bandwidth, with adjustable parameters for context length, batch size and precision (FP16, INT8, Q4). Model metadata pulls from the Hugging Face ecosystem, and you can filter the catalog by use case (coding, long-context, vision), license and commercial usability. A community section invites users to submit real-world benchmark numbers so the estimates get grounded over time.

Editor's take

A genuinely useful niche utility — the kind of thing you would otherwise build in a spreadsheet. The stack-planner mode, which lays out LLM + embedder + reranker across multiple cards, is the standout: it maps directly to how real self-hosted RAG deployments are shaped. Treat the numbers as first-pass sizing, not a benchmark.

— The AI Tool Bible editorial team

Pros

✅ Bilingual Korean/English UI, rare in the self-hosting tools space
✅ Handles multi-GPU stack planning, not just single-model sizing
✅ Precision-aware (FP16 / Q8 / Q4) so quantised deployments get realistic estimates
✅ Covers the full RAG stack: LLM, embedding, reranker, OCR/VLM allocation
✅ Free, no login, runs entirely in the browser
✅ Includes commercial-license filtering for enterprise procurement
✅ Community benchmark submissions ground the theoretical numbers

Cons

⚠️ Estimates are approximations — real throughput depends on driver, kernel and framework specifics not captured here
⚠️ GitHub Pages hosting means no SLA, no accounts and no saved projects
⚠️ Model catalog is limited to what the maintainer curates from Hugging Face
⚠️ No cost modelling versus cloud API alternatives
⚠️ UI is functional but visually spartan compared to commercial capacity planners

Use cases

GPU sizing for self-hosted LLMsMulti-GPU RAG stack planningQuantisation trade-off analysisvLLM deployment capacity checksOllama hardware selectionEmbedding + reranker co-location planningCommercial license filtering for open modelsPre-procurement hardware estimates