Benchmarks and device evidence

This page records what has actually been measured on a real ternux setup, what was only observed, and what remains unverified. It exists to prevent build success, renderer detection and benchmark performance from being mixed into one claim.

Snapshot: evidence supplied by the project author in August 2026. This is one device/configuration, not a compatibility or performance guarantee.

Result classes

A build configured with a GPU option belongs under observed until a runtime log names the device/backend and completes a benchmark or workload.

Tested configuration

The setup notes describe the route as Mesa Zink over Turnip/KGSL. The independent renderer strings from glmark2 and Blender are consistent with that route and do not contain llvmpipe.

Measured: glmark2

Summary

The first command reported OpenGL 4.6 compatibility profile. The second reported OpenGL ES 3.2, but also printed:

MESA-EGL: warning: DRI3 error: Could not get DRI3 device
MESA-EGL: warning: Ensure your X server supports DRI3 to get accelerated rendering

Its renderer still named Zink, Adreno 825 and MESA_TURNIP, and its header still said Surface Size: 800x600 windowed despite the --off-screen command. All three facts are retained. The result is evidence, but it is caveated evidence.

Complete captured scene FPS

GL FPS is the glmark2 run. ES FPS is the separately captured glmark2-es2 --off-screen run. They are placed together only to archive the data; they are not an apples-to-apples comparison.

What the score means

The glmark2 score aggregates scene results. It is not a universal GPU rating and is not the FPS of Xfce, Blender or a game. Resolution, window/fullscreen/off-screen mode, OpenGL versus OpenGL ES, compositor, VSync, Mesa version, thermal state and background activity all change the result.

Only compare runs with the same:

1. glmark2 version and command;
2. API and scene list;
3. resolution and presentation mode;
4. Mesa/renderer stack;
5. compositor and VSync environment;
6. thermal and power conditions.

Observed: Blender 4.3.2

The supplied Blender system report contains:

platform: 'Linux-6.17.0-PRoot-Distro-aarch64-with-glibc2.41'
windowing environment: 'X11'
renderer: 'zink Vulkan 1.4(Adreno (TM) 825 (MESA_TURNIP))'
vendor: 'Mesa'
version: '4.6 (Core Profile) Mesa 26.2.0-devel (git-9452d1daec)'
device type: 'SOFTWARE'
backend type: 'OPENGL'

It also reports Python 3.13.5 and identifies Blender 4.3.2. The Cycles section lists CPU capabilities but no GPU device.

Supported conclusion

Blender launched on X11 and its OpenGL viewport saw the Zink/Turnip renderer.

Unsupported conclusions

This report does not establish:

• Cycles GPU Compute support;
• Blender viewport FPS;
• Eevee or Cycles render time;
• heavy-scene stability or memory limits.

Blender supports Cycles GPU rendering only through its listed compute backends and devices. An accelerated OpenGL viewport is a different path.

Observed: llama.cpp Vulkan build/use

The setup notes report a Debian build configured with:

cmake -B build -DGGML_VULKAN=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release

They also contain local CLI/server examples using small 1–1.5B Q4 GGUF models and GPU-layer offload. They do not contain a complete llama-bench table, build commit, backend/device log, memory reading or thermal record.

Use this for publishable evidence:

cd ~/llama.cpp
./build/bin/llama-cli --list-devices
./build/bin/llama-bench -m models/YOUR_MODEL.gguf -ngl 99 \
  2>&1 | tee ~/benchmarks/llama-bench.txt
git rev-parse HEAD | tee ~/benchmarks/llama-commit.txt

Record model, quantization, model size, context, threads, GPU layers, pp and tg rows, runtime backend/device, RAM and temperature. Until then, ternux makes no llama.cpp tokens-per-second claim for the tested phone.

Observed: stable-diffusion.cpp Vulkan build

The supplied notes report a successful build on the same phone with:

cmake .. -DSD_VULKAN=ON -DCMAKE_BUILD_TYPE=Release
cmake --build . --config Release -j8

No generation command, verbose Vulkan device log, completed image, seed, model, steps, dimensions, timing or memory figure was supplied. This proves a reported Vulkan-enabled build, not measured inference.

A future result should include:

• exact stable-diffusion.cpp commit;
• model and weight format/quantization;
• verbose log naming Vulkan and Adreno/Turnip;
• prompt, negative prompt, seed, sampler, steps and resolution;
• wall time and peak memory;
• output image checksum;
• temperature/power conditions.

Reproduce the graphics measurements

Start the ternux desktop and run inside Debian:

sudo apt update && sudo apt install -y glmark2
mkdir -p ~/benchmarks

date -Is | tee ~/benchmarks/run-context.txt
uname -a | tee -a ~/benchmarks/run-context.txt
glxinfo -B 2>&1 | tee ~/benchmarks/glxinfo.txt
vulkaninfo --summary 2>&1 | tee ~/benchmarks/vulkan-summary.txt

glmark2 2>&1 | tee ~/benchmarks/glmark2-windowed.txt
glmark2-es2 --off-screen 2>&1 | tee ~/benchmarks/glmark2-es2-offscreen.txt

Also record Android/device context from Termux:

ternux profile show | tee ~/ternux-profile.txt
getprop ro.product.model | tee -a ~/ternux-profile.txt
getprop ro.build.version.release | tee -a ~/ternux-profile.txt

For a fair comparison, cool the phone first, keep charger state and screen settings constant, close unrelated apps, state whether Xfce compositing is enabled, and run at least three repetitions. Publish every repetition rather than only the best one.

Quick health benchmark versus evidence reproduction

ternux benchmark is designed as a convenient installation health check. It may try fullscreen glmark2 and uses a timeout, so it is not automatically identical to the two commands archived on this page.

Use:

ternux benchmark
ternux benchmark --json

for quick diagnostics. Use the exact raw commands above when trying to reproduce or compare the published Redmi Turbo 4 Pro evidence.

Current evidence matrix

References

• glmark2
• Mesa Zink
• Blender 4.3 GPU rendering
• llama.cpp Vulkan build
• stable-diffusion.cpp Vulkan build

ternux — Copyright (c) 2026 Sobuj Miah (@soobujmiah) · Apache-2.0