English
Deutsch
PolskiCalculate memory components, context and headroom separately. A sufficient budget establishes neither backend support nor speed.
Calculation and file import run in the browser. These tools do not transmit inputs. Existing website measurement records tool identity and button labels.
Defaults are freely chosen examples. Results are planning estimates.
Import Ollama metadata
Sources and limits
Weights = total parameters × effective bits / 8. Dense KV cache = layers × KV heads × head dimension × tokens × sequences × (K bytes + V bytes). Artifact size is not guaranteed runtime usage. Quantized caches, sliding windows, hybrid and recurrent architectures require suitable models or a measured manual budget. Card profiles show manufacturer GB; the GiB budget is an editable planning assumption and should come from the runtime. Multiple cards are not added together.
- Ollama metadata
- llama.cpp cache types
- NVIDIA RTX 5090
- NVIDIA: RTX 3060-Familie
- NVIDIA: RTX 5060-Familie
- NVIDIA: RTX 3090
- NVIDIA: RTX 5090 32 GB
- NVIDIA: RTX 6000 Ada 48 GB
- NVIDIA: RTX PRO 6000 Blackwell 96 GB
- NVIDIA: RTX 4090 24 GB
- Intel: Arc B580 12 GB / B570 10 GB
- Intel: Arc Pro B50 16 GB
- Intel: Arc Pro B60 und B50
- Intel: Einführung Arc Pro B65 / B70
- Intel: Arc Pro B70 – Datenblatt, 32 GB
- Intel: Arc Pro B65 32 GB – Datenblatt
- Intel: Arc Pro B60 24 GB
Documentation checked: 05 October 2026. Hardware profiles: manufacturer specifications and editable memory budgets.
Questions and answers
What limits apply to the result?
Weights = total parameters × effective bits / 8. Dense KV cache = layers × KV heads × head dimension × tokens × sequences × (K bytes + V bytes). Artifact size is not guaranteed runtime usage. Quantized caches, sliding windows, hybrid and recurrent architectures require suitable models or a measured manual budget. Card profiles show manufacturer GB; the GiB budget is an editable planning assumption and should come from the runtime. Multiple cards are not added together.
How are inputs processed?
Calculation and file import run in the browser. These tools do not transmit inputs. Existing website measurement records tool identity and button labels.