Performance

#The numbers

¹ Average since launch, measured with ps. Arch Linux, Hyprland, GTX 1050 Ti, a 1280×800 window, ReleaseFast build.

² Not measured by us and not a side-by-side test. It's an early Windows 11 build, from a Task Manager screenshot posted by Discord's CTO (coverage). The app isn't out yet, and that screenshot says little about normal use.

³ 125–140 MB in one account's normal use on the same PC, with an i5-2400S, and about 70 MB right after opening. Yours depends on your servers, members and images. Settings → Performance shows where it goes.

⁴ Measured across all 11 themes with demo data, a real GPU, a real window on a hidden Hyprland workspace and GIF autoplay off. A window you're looking at can measure differently.

Read these with the right caveats:

• A real account uses more than demo data. More servers, more members, more images.
• Idle means nothing is moving. An animated GIF on screen costs about 2.5% CPU. The demo's first channel has one, so the measurement turns autoplay off.
• A good part of PSS isn't the app. It's the NVIDIA driver and system libraries such as libcurl and OpenSSL. In one measurement across the themes, anonymous memory was about 25 MB.
• The published Linux file carries SDL3, libcurl and OpenSSL inside, so it runs on more distros. Built against the system's copies it's about 7 MB. These RAM figures come from that kind of build; side by side in demo, the published file used about 2 MB more.
• The Vesktop figures come from a single earlier run on the same machine.

#Measure it yourself

tools/membench.sh 60

For each theme, it opens the app in demo mode with clean settings and GIF autoplay off, waits, measures idle CPU over 5 seconds, reads PSS from /proc/<pid>/smaps_rollup, and fails if any theme goes over the budget in MB. Because of that variation, terminal sometimes fails at 60.

The Performance page in settings shows live numbers, including where the memory goes on your account, and --stats shows the time to first frame.

#Ten decisions

1. It sleeps when nothing happens. The loop waits in SDL_WaitEvent. Network, inotify and the clock wake the app through events. No animation, no redraw. A frame that comes out the same as the last one never reaches the GPU, and a mouse just moving asks for fewer frames.
2. No browser. Everything is drawn with Dear ImGui's ImDrawList on SDL_Renderer.
3. Fonts are mapped, not copied. Each font file is one mmap used by every weight and family, so only the pages read stay in RAM. Loading 4 weights of Inter plus emoji the naive way cost about 90 MB.
4. Only the video driver it needs. On NVIDIA-only machines, glvnd also loaded Mesa and LLVM (about 23 MB) for nothing. The app detects that and avoids it.
5. Markdown is parsed once, when the message arrives. The UI only lays out spans, with a height cache per width.
6. A virtualized list. Binary search for the first visible item, then draw only the screen.
7. Memory has ceilings. 1,500 messages per channel and 3,000 in total, after which the least recently seen conversation is unloaded. Images have a fixed 64 MB budget, and any single download stops at 32 MB. After a minute in the tray or minimized, images leave memory too, except small ones that were on screen.
8. fontconfig gets freed after fonts are resolved (FcFini and malloc_trim).
9. No blinking cursor. It would keep the CPU awake for nothing.
10. Things start when you ask. The voice thread on the first call and FFmpeg on the first video. Off-screen video is released after 30 seconds, and with no plugin on there's no plugin thread, no Lua state and no extra work per message, key or frame.

#The budget in CI

The test suite covers the pure parts, and a separate job builds the whole app and runs the full QA off-screen. See Contributing.

Source of this page on GitHub