lshw: the test machines
Hypoland is measured on real machines, each with Omarchy installed and its full session running, reached over ssh from one desktop. The oldest one is the target; the others make sure that what helps the X200 costs nothing on the GPUs and drivers people actually have.
| Machine | Year | CPU | GPU, Mesa driver | Screen | RAM | What it stands for |
|---|---|---|---|---|---|---|
| Lenovo ThinkPad X200 | 2008 | Core 2 Duo P8400, 2 cores | Intel GMA 4500MHD (GM45, Gen 4.5)crocus, GLES 2.0 only | 1280x800 | 6 GB | The target and the main test machine: stock Hyprland does not start on it. |
| Desktop, AMD A8-6600K APU | 2013 | A8-6600K, 4 cores | AMD Radeon HD 8570D (TeraScale 3)r600, GLES 3.1 | 1920x1200 | 8 GB | The AMD machine: the r600 driver with legacy modesetting. |
| Samsung Series 7 Ultra (NP730U3E) | 2013 | Core i5-3337U, 4 threads | Intel HD Graphics 4000 (Ivy Bridge)crocus, GLES 3.0 | 1920x1080 | 4 GB | Joined on 2026-10-31, Ivy Bridge: the crocus generation between Bay Trail and Haswell. |
| Apple Mac mini Late 2014 | 2014 | Core i5-4260U, 4 threads | Intel HD Graphics 5000 (Haswell GT3)crocus, GLES 3.0 | 1920x1200 over HDMI | 4 GB | Haswell, the last crocus generation: Mesa copies shm windows through a tiled temporary here. |
| Dell Chromebook 11 3120 | 2015 | Celeron N2840, 2 cores | Intel Bay Trailcrocus, GLES 3.0 | 1366x768 | 4 GB, 16 GB eMMC | The low end with GLES 3: two slow cores, little memory, wifi only. |
| Lenovo ThinkPad T550 | 2015 | Core i5-5300U, 4 threads | Intel HD Graphics 5500 (Broadwell)iris, GLES 3.2 | 1920x1080 | 16 GB | The second machine, a plain iris laptop; the comparison with stock Hyprland on GLES 3 hardware started here. |
| Dell XPS 13 9343 | 2015 | Core i5-5300U, 4 threads | Intel HD Graphics 5500 (Broadwell)iris, GLES 3.2 | 3200x1800 at scale 2 | 8 GB | The high-resolution one: four times the pixels of the X200 on the same class of GPU. |
| Lenovo ThinkPad T460s | 2016 | Core i7-6600U, 4 threads | Intel HD Graphics 520 (Skylake)iris, GLES 3.2 | 1920x1080 at scale 1.5 | 20 GB | The fractional scale, and the only Skylake: its scanout buffers are render compressed. |
| Lenovo ThinkPad T490 | 2019 | Core i5-8265U, 8 threads | Intel UHD Graphics 620 (Whiskey Lake)iris, GLES 3.2 | 1920x1080 | 8 GB | Joined on 2026-10-30, the newest Intel GPU of the lab. |
| Apple Mac mini M1 | 2020 | Apple M1, 8 cores, aarch64 | Apple M1 GPU (G13)asahi, GLES 3.2 | 1920x1080 over HDMI | 8 GB | Apple Silicon with the Asahi kernel and Mesa, joined on 2026-11-10; no cursor plane, and its display driver reports no page flips to count. |
| Samsung Galaxy Book4 Edge 14" | 2024 | Snapdragon X Elite, 12 cores, aarch64 | Qualcomm Adreno X1-85freedreno, GLES 3.2 | 2880x1800 at scale 2, 120 Hz | 16 GB | The aarch64 machine, with its own package build; 120 Hz. |
An HP EliteBook 8560p (2011, Radeon HD 6470M, r600) was in the lab from 2026-10-22 until its GPU failed on 2026-10-27. Not in the lab: NVIDIA
(nouveau), AMD GCN and newer (radeonsi, amdgpu), and the two other Intel
generations that stop at GLES 2.0, GMA X3100 (965GM) and Ironlake. Reports from any of them are welcome.
bench: Hypoland against Hyprland 0.57.1, per machine
| Hypoland / Hyprland 0.57.1 | GPU client at 60 fps | foot terminal scrolling fast | Chromium in a full window | workspace switching | memory (PSS), 60 s after the start | frames on the screen, GTK 3 scrolling | a focus change shown |
|---|---|---|---|---|---|---|---|
| Lenovo ThinkPad X200 | 2.6% / 4.2% | 9.0% / 14.1% | 8.5% / 14.7% | 4.1% / 7.9% | 42 MiB / 116 MiB | 60 fps / 31 fps | 4.9 ms / 5.4 ms |
| Desktop, AMD A8-6600K APU | 3.4% / 3.8% | 2.6% / 14.4% | 3.3% / 4.2% | 3.2% / 7.7% | 42 MiB / 115 MiB | 60 fps / 59 fps | 1.7 ms / 2.0 ms |
| Samsung Series 7 Ultra (NP730U3E) | 3.2% / 6.4% | 4.3% / 15.9% | 4.6% / 6.0% | 4.0% / 7.8% | 42 MiB / 117 MiB | 46 fps / 34 fps | 2.6 ms / 14.9 ms |
| Dell Chromebook 11 3120 | 6.1% / 9.2% | 5.2% / 15.1% | 7.0% / 9.0% | 5.6% / 10.9% | 43 MiB / 115 MiB | 60 fps / 60 fps | 7.8 ms / 8.4 ms |
| Lenovo ThinkPad T550 | 3.7% / 4.5% | 3.7% / 13.4% | 3.9% / 5.4% | 3.7% / 7.1% | 43 MiB / 118 MiB | 60 fps / 60 fps | 2.3 ms / 7.8 ms |
| Dell XPS 13 9343 | 4.0% / 5.8% | 3.9% / 22.8% | 4.4% / 6.5% | 4.0% / 8.3% | 43 MiB / 118 MiB | 44 fps / 40 fps | 2.5 ms / 12.5 ms |
| Lenovo ThinkPad T460s | 4.1% / 4.7% | 3.2% / 9.2% | 4.2% / 5.2% | 3.2% / 6.4% | 43 MiB / 119 MiB | 60 fps / 60 fps | 2.0 ms / 12.0 ms |
| Lenovo ThinkPad T490 | 3.4% / 4.2% | 4.2% / 15.1% | 4.2% / 4.9% | 3.0% / 5.2% | 43 MiB / 119 MiB | 60 fps / 59 fps | 1.7 ms / 8.1 ms |
| Apple Mac mini M1 | 0.9% / 0.9% | 4.5% / 5.5% | 1.4% / 2.3% | 1.0% / 2.2% | 49 MiB / 124 MiB | – | 0.8 ms / 1.0 ms |
| Samsung Galaxy Book4 Edge 14" | 4.1% / 4.6% | 6.6% / 7.3% | 2.9% / 3.5% | 3.6% / 3.8% | 43 MiB / 146 MiB | 120 fps / 120 fps | 1.3 ms / 4.8 ms |
Compositor CPU in percent of one core, medians of 8 to 10 fresh sessions of each compositor per machine,
alternating from 2026-11-10 23:09 to 11-11 07:30, Hypoland master r957 (after 0.3.0) against Hyprland 0.57.1-2. Apple Mac mini Late 2014: not in the table, it had no monitor in the night of the baseline (its HDMI went to a KVM), so its numbers are left out. Both ran without realtime scheduling, which Omarchy turns off for Hyprland now
(HYPRLAND_NO_RT=1), except on the Galaxy Book. Green where
Hypoland is at least 5% better, red where it is at least 5% worse. On the X200, Hyprland ran with Omarchy's
GLES 3 override (omarchy#14764), since it does
not start there otherwise. Hyprland ran with quirks:skip_non_kms_dmabuf_formats = false,
Hypoland's default: with its own default the T550's session came up without a wallpaper and GL clients that ask
for alpha got no config. The frames are the compositor's own page flips; the Galaxy Book's screen runs at
120 Hz, and the Mac mini M1's display driver reports none. The memory holds the shell's shared pages too.
Four terminals side by side on every machine: terminals; games and 3D
benchmarks: games; three file managers: files.
Raw data, every session's values and the scripts.
measure.sh: how it is measured
Every number comes from a fresh session: the machine's display manager is restarted, the desktop settles for a minute, then the benchmark runs its workloads with the compositor's real clients (foot, Alacritty, Chromium, GTK 3 and 4, Qt Quick, Tk through Xwayland, Omarchy's shell). Keys and pointer motions come from a uinput device, the same path as a real keyboard. Latencies are the time from an input event to the first screen-capture frame that shows its effect. Two builds alternate through a night, so a drift of the machine hits both the same.
perf: what else is looked at
- GPU time per frame, by the compositor's own pass profile
- Power, by RAPL or the battery's discharge rate, net of an idle baseline
- Memory: PSS, heap, threads and the GPU buffers the compositor holds
- Profiles with
perfand the kernel's and Mesa's symbols - Every machine's journal and kernel log stream to the desktop, so a machine that dies leaves its last words
ls docs/baseline: the data
- docs/baseline: every comparison with stock Hyprland, with the raw values and the scripts that made them
- scripts/bench: the benchmark, its workloads and clients
- AGENTS.md: the engineering notes, with what was tried, measured and dropped
- CHANGELOG.md: the changes per release with their measured effect