LLM Hardware team
Updated
The editorial team at LLMHardware.io. Researches published specs and model cards for running local LLMs on consumer hardware.
Contact: contact page
About the team
The LLM Hardware team writes and maintains LLMHardware.io. Pages are researched from published specs, vendor spec sheets and model cards, and sized with the site's calculator and VRAM formula. Drafting is AI-assisted, with an edit pass before publishing.
There is no lab and no vendor sample hardware behind this byline. The method is to cite published llama.cpp llama-bench runs, the XiongjieDai GPU-Benchmarks-on-LLM-Inference repository, the Home GPU LLM Leaderboard, Hardware Corner's GPU ranking, and the VRAM formula. A speed figure names its source. If it has no source, it is not on the page.
The full methodology, including the VRAM formula, the worked example, and the source list, is on the methodology page.
How the numbers are produced
The VRAM figures on the guides use one formula: params × bytes × 1.2 + 2. Q4 is 0.5 bytes per parameter, so a 32B Q4 estimate is 21.2 GB. Speed numbers, when a guide has them, come from published community runs: llama.cpp llama-bench, the XiongjieDai GPU-Benchmarks-on-LLM-Inference repository, the Home GPU LLM Leaderboard, and Hardware Corner. This site does not measure hardware.
What we cover
- Consumer GPU comparisons for local inference (NVIDIA RTX, AMD Radeon, Intel Arc)
- Apple Silicon for LLMs (M3 and M4 series, unified memory tradeoffs)
- Multi-GPU setups for larger models, including NVLink and CPU offloading
- Published power and thermal figures for running models around the clock
- Quantization in practice: GGUF formats, Q4 vs Q5 vs Q8 quality drops as reported by published evals
- Setup notes for Ollama, llama.cpp, LM Studio, vLLM, and Open WebUI on consumer machines
How we cite
Fit figures ("fits in N GB of VRAM") come from the site VRAM formula. Speed figures on some guides are approximate and not all are cited yet. The full source list and the VRAM formula live on the methodology page. The short version: llama.cpp's own llama-bench discussion thread for Apple Silicon baselines, the XiongjieDai community repo for NVIDIA + Apple Silicon llama-bench runs, the Home GPU LLM Leaderboard for VRAM-tier comparisons, and Hardware Corner for context-length sensitivity.
If you spot a number that looks wrong, we want to hear about it.
What we will not claim to be
- A benchmark publisher. The site does not publish its own benchmark results. If a claim looks wrong, please use the contact page.
- An academic researcher. This site does not publish papers or controlled studies of model quality.
- A training specialist. Fine-tuning at scale is outside these pages. The subject is inference memory and the hardware that holds it.
- A datacenter engineer. Recommendations on this site are for home labs, small offices, and individuals, not racks of H100s.
- A model-evaluation lab. When a page prefers one model for a task, it says which memory tier it fits and links the card. It does not report a private score.
Contact and corrections
Spot something wrong in a guide? Use the contact page with the page URL and the correction. Confirmed errors are logged on the corrections page and the original guide is updated with a dated note.