CPU offload split estimator

One-sentence answer: Apply a simple penalty when part of the model is offloaded to host memory.

Educational calculator. Formulas are plain text. Replace defaults with your documents when accuracy matters.

Full-GPU tok/s

Offload fraction 0–1

Offload factor 0–1

How it is calculated

effective ≈ tps × (1 − frac + frac × penalty_factor).

Worked example

40 tok/s, 25% offload, factor 0.15 → about 31.5 tok/s.

Limits

Not a substitute for official quotes, tax software, or licensed advice.
Local rules and product terms can differ from this simplified model.
Static page: inputs are not stored on a server.

Embed:

<iframe src="https://vramwise.com/tools/offload-split/?embed=1" title="CPU offload split estimator" loading="lazy" style="width:100%;min-height:420px;border:0"></iframe>
<p><a href="https://vramwise.com/tools/offload-split/">CPU offload split estimator by VRAMwise</a></p>

Offload guide

All tools · Guides · Methodology

Disclaimer

Educational use only. Not tax, legal, investment, or purchasing advice.

Field guides

context budgeting · cpu offload speed · how we calc vram · kv cache math

Catalog exits

Tools · Guides · Data

Frequently asked questions

What is this for?

Sketch speed loss when a fraction of layers run off-GPU. Use the formulas and examples; change one input at a time.

Is this advice?

No. Educational estimates only. Confirm with primary sources and professionals.

How fresh are defaults?

Defaults cite as-of dates on data pages or tool notes.

Can I embed it?

Yes — keep the attribution link on the parent page.

Any AdSense code?

No ad script loader or shared pub-id is embedded here.

Result looks wrong?

Check units and model limits, then contact with URL and inputs.