# Python GUI benchmarks — Pyzed vs tkinter, Qt, Kivy, Flet… > Measured, reproducible benchmarks of Python GUI toolkits: import time, time to first frame, memory, update latency with 1,000 rows, install size and lines of code for Pyzed, tkinter, CustomTkinter, PySide6, Kivy, Dear PyGui, Flet and Toga. Source: https://albakore.github.io/pyzed/why/benchmarks/ These numbers were **measured**, not estimated: every chart on this page is drawn from the JSON file the benchmark script wrote, and you can run the same script on your machine. They come from one machine without a GPU, so read them as a comparison between toolkits under the same conditions, not as absolute numbers for your computer. :::caution[Software rendering] The machine has no GPU. Windows were shown on a virtual X server (Xvfb) and every toolkit that uses OpenGL or Vulkan (Pyzed, Kivy, Dear PyGui, Flet) rendered on the CPU through Mesa's llvmpipe and lavapipe drivers. That is the worst case for GPU toolkits and the normal case for tkinter, Qt Widgets and GTK, which draw on the CPU anyway. On a computer with a GPU, the drawing part of the GPU toolkits' numbers moves off the CPU. ::: ## Key findings - **Startup is fast everywhere except Flet and, with many widgets, Kivy and Toga.** Pyzed shows a 10-row window in about 0.16 s from launch and a 1,000-row window in about 0.24 s: building and laying out 2,000 more widgets costs it some 80 ms. tkinter, Dear PyGui and Qt start sooner; Kivy, Toga and Flet need close to a second for 1,000 rows. - **Updating 1,000 rows takes Pyzed about 18 ms** from state change to screen, even though it rebuilds the whole view from Python each time. That is a little faster than Qt Widgets, about 3× faster than Kivy and 9× faster than Flet; tkinter and Dear PyGui are faster still. Toga (GTK) needed about 37 s per update in this test, because every label change triggers a new layout. - **Memory depends on the renderer.** Through the software Vulkan driver used here, Pyzed's GPU renderer takes about 151 MB; the same app on Pyzed's CPU renderer (tiny-skia, set with `ICED_BACKEND=tiny-skia`) takes about 48 MB, between tkinter (25 MB) and Qt (72 MB). The other GPU toolkits pay a similar driver cost here (Dear PyGui 123 MB, Kivy 146 MB, Flet 268 MB). - **Install size:** about 30 MB for Pyzed, against 238 MB for Qt's essentials, 83 MB for Kivy and 51 MB for Flet with its client. tkinter, Dear PyGui, Toga (which uses the system's GTK) and CustomTkinter (which uses tkinter) are smaller. - **Code:** the to-do app takes 31 lines in Pyzed, half of PySide6 (62) and well under Kivy (81). Flet's is the shortest (22). ## Startup **Import** is the time to `import` the toolkit in a fresh Python process. **First frame** is the time from launching the app's process until its window is on the screen with its content: the script watches the screen for a colored bar at the bottom of the window. ## Memory Resident memory (RSS) of the app's processes one second after the first frame. For Flet it includes the Flutter client process that draws the window. ## Updating 1,000 rows Every 250 ms the app changes the text of all 1,000 labels and the color of the bottom bar in the same update. The latency is the time from the app changing its state to the new color being on the screen; 10 updates per launch, 7 launches. ## Install size How much a fresh virtual environment grows when the toolkit is installed with its dependencies. tkinter counts the Tcl/Tk files that come with Python. Flet includes the desktop client it downloads on first run. Toga does not include GTK, which comes from the operating system. NiceGUI is shown for reference only: it draws in a web browser, so it was not part of the window benchmarks. ## Lines of code The same counter and to-do apps written in each toolkit ([source](https://github.com/albakore/pyzed/tree/main/benchmarks/apps)), counting lines that are not blank, comments or docstrings. See [Developer experience](https://albakore.github.io/pyzed/why/developer-experience/#how-much-code) for the charts and what the lines do. ## Raw numbers Median in bold; below it the minimum–maximum and the number of samples. The complete results, with every sample, are in [`benchmarks/results/`](https://github.com/albakore/pyzed/tree/main/benchmarks/results). ## Machine and versions ## Method - **Same app everywhere.** Each toolkit's `bench.py` builds a 640×480 window with a scrollable list of N rows (a label and an "Open" button each) and a 40 px bar at the bottom painted in a sentinel color. Lists are built the plain way each toolkit documents, with one widget per row: no `RecycleView` in Kivy, no lazy `ListView` in Flet, no `virtual_list` in Pyzed. - **First frame** is measured from outside: the harness starts the process and grabs the screen about every millisecond (with [mss](https://github.com/BoboTiG/python-mss)) until 8,000 pixels of the sentinel color are visible. - **Memory** is the RSS of the process and its children ([psutil](https://github.com/giampaolo/psutil)), one second after the first frame. - **Updates** are timed between the app's own timestamp (`time.monotonic()`, printed right before it changes its state) and the moment the harness sees the bar's new color. - **Repetitions:** 7 launches per toolkit and size, 10 updates per launch, 10 imports; the charts show the median and a thin line from the minimum to the maximum. - **Pyzed (CPU renderer)** is the same Pyzed app started with `ICED_BACKEND=tiny-skia`, which makes Iced use its CPU renderer instead of wgpu: what Pyzed falls back to on machines without a working GPU driver. - **Install size:** a fresh `uv venv`, `uv pip install --link-mode=copy `, size difference. ## Reproduce it Without a display it starts Xvfb by itself; on a desktop, keep the benchmark windows uncovered while it runs. See [`benchmarks/README.md`](https://github.com/albakore/pyzed/blob/main/benchmarks/README.md) for the options and the system packages some toolkits need. Results on other machines are very welcome in an [issue](https://github.com/albakore/pyzed/issues). ## Caveats - **One machine, no GPU** (see above). Results on Windows and macOS, and with a real GPU, will differ. - **The harness shares the CPU** with the app: grabbing the screen costs about one core while it polls. - **Tk draws widget by widget**, so for tkinter the bar may appear a moment before every row has been drawn; the GPU toolkits present whole frames. - **Dear PyGui redraws continuously**, so its update latency includes waiting for its next frame; it also uses the CPU while idle, which is not measured here. - **CustomTkinter has no 1,000-row results**: its app died under Xvfb with an X error (BadAlloc or BadValue from `CreatePixmap`) once its scrollable frame held more than about 100 rows. - **Not measured:** wxPython has no Linux wheels on PyPI and building it from source was not practical here; NiceGUI runs in a web browser, so a desktop window timing would measure the browser. - **Lines of code** depend on style; the apps follow each toolkit's documented idioms, and improvements are welcome.