> ## Documentation Index
> Fetch the complete documentation index at: https://docs.echophrase.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Benchmarks

> Measured resource usage - install size, RAM, GPU power, and VRAM - with full methodology

# Benchmarks

Echophrase is built on **Tauri + Rust** instead of Electron, and we measured
exactly what that buys you. These are first-party numbers, but they're honest
ones: where a figure is an estimate or a comparison is indirect, we say so.

## TL;DR

Measured 2026-07-06 on an NVIDIA GeForce RTX 5080 with **NVIDIA Parakeet-TDT
0.6B** (the default model) loaded:

| Phase                                   | GPU Power  | GPU VRAM    | RAM         | CPU     |
| --------------------------------------- | ---------- | ----------- | ----------- | ------- |
| GPU idle (app closed)                   | \~21.9 W   | 2690 MiB    | —           | —       |
| App + model loaded, ready               | \~35.6 W   | 5217 MiB    | 226.6 MB    | 0%      |
| Transcribing one phrase (marginal peak) | **+7.8 W** | **+31 MiB** | **+4.3 MB** | 5% peak |

A single transcription is a marginal blip - a fraction of a watt-second and a
few megabytes - then everything settles back to the ready state.

## Install size

The extracted Windows app is **34.2 MB**. For context: any Electron-based app
carries a **\~200 MB runtime** (bundled Chromium + Node + V8) before a single
line of app code is added - Echophrase's entire install is roughly **6x
smaller than that empty floor**. Real Electron apps commonly land at
300-500 MB installed.

Sources for the Electron runtime figures (independent 2026 framework
comparisons, not our own measurements):
[Tech Insider](https://tech-insider.org/tauri-vs-electron-2026/),
[Hopp](https://www.gethopp.app/blog/tauri-vs-electron),
[codenote.net](https://codenote.net/en/posts/famous-electron-apps-2026-research/).

## Memory behavior: not a leak

RAM and VRAM rise after the first couple of transcriptions (226.6 → 390.7 MB
RAM), which can look leak-like. We stress-tested it: after \~5 more
back-to-back utterances with no restart, RAM moved from 385.31 MB to just
385.48 MB and VRAM held flat.

|                           | RAM       | GPU VRAM   |
| ------------------------- | --------- | ---------- |
| Before the batch          | 385.31 MB | \~5343 MiB |
| After \~5 more utterances | 385.48 MB | 5332 MiB   |

The initial rise is a **one-time warm-up buffer/cache** that plateaus and
stays flat regardless of how many phrases follow - a sustainable, always-ready
footprint rather than one that degrades over a session.

## Methodology

<Note>
  **Hardware:** NVIDIA GeForce RTX 5080. **Model:** NVIDIA Parakeet-TDT 0.6B.
  GPU power/VRAM/utilization sampled via one long-lived
  `nvidia-smi --query-gpu ... -lms` process; RAM/CPU via PowerShell
  `Get-Process`, with %CPU computed Task-Manager-style (delta CPU-seconds ÷
  delta wall-seconds ÷ logical cores). Sampler overhead is small but not zero -
  treat these as close estimates, not lab-isolated numbers. The GPU idle floor
  also drifts between sessions with unrelated system activity, so each run
  captures a fresh baseline.
</Note>

## How this compares

To our knowledge, **no other app in this category publishes a controlled
GPU-watt or per-utterance VRAM benchmark**, so there is no like-for-like
public comparison for those numbers - we simply measured what nobody else
reports. Where independent data does exist (the Electron runtime and install
sizes above), we cite it; where it doesn't, we say so rather than overclaim.

<Tip>
  Questions about the numbers, or want them run on different hardware?
  [Join the Discord](https://discord.com/invite/Er6NUcx).
</Tip>