AI Hardware Builder
Set a budget. Get a complete local-AI rig.
Like PCPartPicker, built for people running models at home. Tell us your budget and what you want to do — we pick the GPU, CPU, PSU, RAM and SSD that fit, then estimate real speed, context, running cost and where to upgrade next.
—
—
—| Part | Component | Price |
|---|---|---|
| Total build | — |
—
We put most of the budget into the GPU — it decides everything for AI — then right-size the CPU (inference is GPU-bound, so no need to overspend), a PSU with real headroom, RAM at roughly 2× your VRAM for offload, and an SSD sized to your workload. Speeds are bandwidth-bound decode ceilings; running cost uses your hours and rate.
Built your rig?
See exactly what it runs.
Take this build to the compatibility scanner and watch every popular model light up green.
Evidence & method
How this calculation works
Sizes a build around the models you want to run — estimating required VRAM and matching GPUs that clear it.
Data sources
- NVIDIA, AMD & Intel GPU documentation
- Hugging Face model cards
- Official model papers
- Bitpute Methodology
Assumptions
- Weights = parameters × bits-per-weight ÷ 8
- Runtime overhead = 0.75 GB + 5% of weights (CUDA context, allocator, buffers)
- Recommended VRAM = total × 1.10 (10% headroom)
- Usable VRAM = card capacity × 0.95
- KV cache counted only when the model architecture is known; when applied, batch size 1, FP16 KV, context as entered
- Single or multiple identical GPUs, inference workload
Limitations
- Actual VRAM varies by framework (PyTorch, llama.cpp, vLLM, TGI), driver/CUDA version and OS
- Quantization implementations differ; real bits-per-weight can vary from the nominal value
- When architecture is unknown the KV cache is omitted, so long-context use will exceed the estimate
- Estimates are for planning and comparison, not a guarantee
Data status: Hardware specs — vendor documentation. Model metadata — community model cards, not independently verified.
Related
Why does this estimate differ from other calculators?
- Different footprint formulas and rounding
- Different KV-cache assumptions (batch, context, precision)
- Framework and inference-engine differences
- Reserved and fragmented VRAM
- Driver and CUDA overhead
- Precision and quantization choices