The honest buyer's check
“Why not?” — every GPU has a ceiling
Most sites only tell you what to buy. This one shows you where a card runs out of room before you spend the money — and the cheapest way past each limit.
Pick a GPU to reality-check
Every ceiling below is computed from real model weights, KV-cache and fine-tuning memory — not a spec sheet. VRAM figures assume 4-bit (Q4) quantization. See the Can I Run It? checker for a specific model.
Evidence & method
How this calculation works
Compares a GPU's usable VRAM and memory bandwidth against a model's footprint to explain whether — and why — it will run well.
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 GPU, 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