Pixel & Oak/Free Tools/Local AI VRAM Calculator
Runs Locally · bandwidth arithmetic
What your card can actually hold and read
Decoding one token means reading every weight, so the speed limit is just
bandwidth ÷ bytes read per token. Pick a machine and a build below.
Every figure here is a manufacturer specification or a file size read off Hugging Face —
these are ceilings, not benchmarks.
Where the conversation runs out
Memory needed climbs as the chat grows — weights are fixed, the cache is not.
Where the climbing line meets your machine's flat limit is the whole answer.
Every machine, same build
Sorted by reading speed. Your selection is highlighted.
What this does not claim
- Ceilings, not benchmarks. Real decoding lands near 60–80% of these numbers depending on backend, batch size and quantisation. Nothing here was measured on hardware.
- Usable memory is less than the sticker. Display output and driver reserve take a share, so each machine's usable figure is set a little below its nominal size and is typical, not guaranteed.
- Decode only. Reading your prompt in the first place (prefill) is compute-bound and is a separate cost.
- Prices move. The used-card figure is a market price, not a manufacturer one.
Sources
- DGX Spark — NVIDIA hardware specifications: 128 GB LPDDR5x, 256-bit, 4266 MHz, 273 GB/s.
- RTX 5090 desktop and Laptop GPU — NVIDIA GPU comparison: 32 GB / 512-bit and 24 GB / 256-bit. Bandwidth derived from bus width at 28 Gbps.
- Ryzen AI Max+ 395 — AMD specifications: 128 GB LPDDR5x, 8000 MT/s, 256 GB/s.
- Model file sizes — ggml-org and unsloth GGUF repositories, read 2026-08-23.
- Cache rate of 64 KiB per token derives from
config.json: 16 full-attention layers of 64 layers total.