Model deployment estimate

LLM / NPU Memory Calculator

Estimate model weight memory, KV Cache, and approximate prefill compute from parameters, batch, sequence length, layers, hidden size, and precision.

Local

Input

INPUT
Assumptions Weight memory = parameters × bytes per parameter; KV Cache = batch × seq × layers × hidden × 2(K/V) × bytes; prefill FLOPs ≈ 2 × parameters × batch × seq.

Ready.

Result

OUTPUT

Weight Memory

KV Cache

Total Memory

Prefill Compute

Total Tokens

Bytes / Param

Formula Detail


        
The LLM / NPU Memory Calculator helps estimate whether a large model can fit on a local GPU, NPU, or edge inference device. It separates model weights from KV Cache and reports MB/GB plus approximate prefill FLOPs so FP32, FP16, INT8, and INT4 deployments can be compared quickly. All calculations run locally in JavaScript.

Practical guide

LLM / NPU Memory Calculator

Estimate whether a model configuration fits an accelerator before you start a run, buy hardware, or change quantization settings.

How to use

  1. Enter the parameter count in billions, the precision used for weights, batch size, sequence length, layer count, and hidden size.
  2. Review the separated weight memory, KV cache memory, total estimate, and rough compute figure.
  3. Change precision or sequence length to compare the pressure from FP16, INT8, INT4, long context, or larger batches.
  4. Copy the result into a hardware note, sizing spreadsheet, or deployment checklist.

Common use cases

  • Checking whether a 7B or 13B model can fit on a local NPU, mobile accelerator, or single GPU.
  • Explaining why long-context inference grows quickly because KV cache scales with batch, layers, sequence length, and hidden size.
  • Comparing FP16, INT8, and INT4 assumptions before building a benchmark or purchasing memory capacity.

Privacy note

All values are numbers typed into the page and all calculations run in browser memory. The tool does not send model names, sizing assumptions, or hardware notes to a server.