================================================================
  Arminius AI Benchmark
  Pure Vulkan Rendering Pipeline Inference
  https://gamedev.tech/benchmark
================================================================


WHAT IS THIS?

  This benchmark runs a real 4-billion parameter AI model (Gemma 3 4B)
  through your GPU's Vulkan rendering pipeline -- the same hardware
  path that renders video games.

  No CUDA. No ROCm. No PyTorch. No ML frameworks.
  One binary. One graphics API. Raw GPU hardware performance.

  Your result is uploaded to the global leaderboard and your browser
  opens directly to your score.


QUICK START

  Option A — Double-click arminius.exe (Windows) or ./arminius (Linux)
    Opens the graphical interface. The benchmark runs automatically.
    Results are uploaded and a shareable result card is generated.

  Option B — Command line via script
    Windows:  run_benchmark.bat
    Linux:    bash run_benchmark.sh
    SteamOS:  bash run_benchmark.sh (Desktop Mode, Konsole)

  The model (~4 GB) is downloaded automatically on first run.


GUI MODE

  Double-clicking arminius opens a graphical interface with:

    Left:  Benchmark results, score, GPU info
    Right: Chat interface (talk to the AI model directly)

  The chat feature is an early preview. It demonstrates that the
  Vulkan rendering pipeline runs real inference, not a pre-computed
  demo. Full chat functionality (system prompts, longer responses,
  conversation history) is under development.


COMMAND LINE USAGE

  Download the model (one-time, ~4 GB):
    arminius --download-model

  Benchmark your GPU:
    arminius --benchmark --tag ABC

  Benchmark with custom model path:
    arminius --benchmark --vindex /path/to/model --tag ABC

  Chat with the model:
    arminius --chat

  Skip upload:
    arminius --benchmark --tag ABC --no-upload


CUDA / PYTORCH COMPARISON (for researchers)

  If you have Python and PyTorch with CUDA installed, you can run a
  side-by-side comparison on the same GPU.

  Requirements:
    - Python 3.8+
    - PyTorch with CUDA:   pip install torch
    - For AMD ROCm:        pip install torch --index-url
                           https://download.pytorch.org/whl/rocm6.2

  Run with CUDA comparison:
    arminius --benchmark --pytorch --tag ABC

  On multi-GPU systems, select the GPU:
    CUDA_VISIBLE_DEVICES=0 arminius --benchmark --pytorch --tag ABC

  The tool will:
    1. Run the Vulkan benchmark (burst + sustained throughput)
    2. Download pytorch-baseline.py from gamedev.tech
    3. Run the same model through PyTorch/CUDA on the same GPU
    4. Display both results side-by-side (Vulkan vs CUDA)
    5. Submit the comparison to the leaderboard

  The PyTorch comparison is optional. The Vulkan benchmark works
  without Python, CUDA, or any ML framework installed.

  Typical results: 1.3x-2.2x faster via Vulkan rendering pipeline
  compared to CUDA/PyTorch on the same hardware.


SYSTEM REQUIREMENTS

  - Any GPU with Vulkan 1.1+ support
  - 4 GB+ VRAM (model uses ~2 GB via SNORM texture compression)
  - ~4 GB disk space for model files
  - Windows 10/11 or Linux (Ubuntu 22.04+, Fedora, Arch, SteamOS)


TESTED GPUs (25+)

  NVIDIA:  RTX 5090, RTX 4090, RTX 3090/Ti, RTX 3060, GTX 1080 Ti,
           TITAN V, TITAN X (Pascal), TITAN RTX,
           Quadro RTX 8000/5000, RTX A6000/A5000/A4000,
           Tesla V100, P100, P40, M40, Quadro M4000
  AMD:     Radeon RX 7900 XTX, Vega 64, Steam Deck (Van Gogh APU)
  Intel:   Arc (via ANV Mesa Vulkan driver)


WHAT DATA IS COLLECTED

  The benchmark uploads only the following technical data:
    - GPU model name, architecture, VRAM size
    - Benchmark scores (tokens/second, latency, power draw)
    - Your 3-letter tag (freely chosen, not linked to identity)
    - Operating system type (Windows/Linux/SteamOS)

  NO personal data is collected, stored, or processed. No IP addresses
  are logged. No cookies are used. No tracking of any kind. No account
  or registration required.

  All submitted data is purely technical benchmark telemetry used to
  display the public leaderboard at gamedev.tech/benchmark.

  Use --no-upload to run the benchmark without submitting any data.

  This data processing is based on legitimate interest (Art. 6(1)(f)
  GDPR/DSGVO) for the purpose of providing a public GPU performance
  comparison service. The data is limited to what is strictly necessary
  for this purpose (data minimisation, Art. 5(1)(c) GDPR/DSGVO).


CONTACT & LEGAL

  Website:     https://gamedev.tech/benchmark
  Discord:     https://discord.gg/HS3WTmutag
  Contact:     info@gamedev.tech

  Website & benchmark service operated by:
  Strategische Spiele Entwicklung UG (haftungsbeschraenkt)
  Geschaeftsfuehrer: Valentin Schierhuber
  Registergericht: Amtsgericht Darmstadt, HRB 104014

  Arminius software & technology (c) 2026 Gregor Koch.
  All rights reserved. Patents pending.
