arcbench

Intel Arc GPU benchmark console for Linux — a guided TUI for running, validating, and sharing GPU benchmark results.

arcbench

Built for Intel Arc GPU owners on Linux who want comparable, shareable results without memorizing CLI flags.

---

Features

• System preflight — detects GPU, oneAPI, llama-bench, and telemetry availability on launch
• Guided benchmark runner — preset configurations matching the Intel-Arc-GPU-Benchmarks repo format
• Live telemetry — GPU VRAM, CPU, RAM monitoring during benchmark runs
• Result validation — checks for SYCL backend, both pp/tg rows, and GPU activity
• Repo-compatible export — Markdown output ready for GitHub contribution

---

Quick start

Prerequisites

• Ubuntu 22.04+ or Debian 12+
• Intel Arc GPU with xe driver
• Intel oneAPI BaseKit (/opt/intel/oneapi/setvars.sh)
• llama-bench built with SYCL backend (see setup guide)

Install

Option A — uv (recommended, fastest)

uv manages its own isolated environment automatically. No manual venv activation needed.

git clone https://github.com/Wesley-Jakob-Gilbert/arcbench
cd arcbench
uv venv              # creates .venv/ in the project directory
source .venv/bin/activate
uv pip install -e .
arcbench

Important: uv pip install installs into whichever environment is currently active. If you run it without activating the venv first, the package lands in your system Python and the arcbench command may not be on your PATH (or may conflict with system packages). Always source .venv/bin/activate before installing or running.

To re-enter the environment in a new shell:

cd arcbench
source .venv/bin/activate
arcbench

Option B — pipx (install once, run anywhere)

pipx creates and manages the venv for you — no activation step ever needed.

pipx install git+https://github.com/Wesley-Jakob-Gilbert/arcbench
arcbench

Option C — standard pip from source

git clone https://github.com/Wesley-Jakob-Gilbert/arcbench
cd arcbench
python3 -m venv .venv
source .venv/bin/activate
pip install -e .
arcbench

---

Benchmark coverage

Machine Learning (llama.cpp SYCL)

Preset	Model	Size	Notes
Llama 3.1 8B Q4_K_M	Meta-Llama-3.1-8B-Instruct	4.58 GiB	Quick sanity + repo standard
Llama 3.1 8B Q5_K_M	Meta-Llama-3.1-8B-Instruct	5.33 GiB	Higher quality
Llama 3.1 8B Q8_0	Meta-Llama-3.1-8B-Instruct	7.95 GiB	Memory bandwidth wall demo
Llama 3.1 70B Q2_K	Meta-Llama-3.1-70B-Instruct	24.56 GiB	Largest all-GPU fit on 32GB
Llama 3.1 70B IQ3_XS	Meta-Llama-3.1-70B-Instruct	27.29 GiB	3.3bpw, all-GPU
Llama 3.1 70B Q4_K_M hybrid	Meta-Llama-3.1-70B-Instruct	39.59 GiB	Hybrid CPU+GPU — ngl 60

Coming soon

• Blender benchmark runner
• GROMACS / HPC workloads
• Multi-GPU configurations

---

Project structure

arcbench/
  arcbench/
    app.py              # Textual app entry point
    screens/
      welcome.py        # System scan
      home.py           # Dashboard + category select
      category.py       # Benchmark list + detail panel
      running.py        # Live benchmark + telemetry
      completion.py     # Results + export
    workers/
      benchmark.py      # llama-bench subprocess management
      telemetry.py      # GPU/CPU/RAM polling
    data/
      system.py         # Hardware/software detection
      manifest.py       # Benchmark preset loader
      results.py        # Local result storage
  benchmarks/
    llama_cpp/
      llama31.yaml      # Llama 3.1 preset definitions

---

Contributing benchmarks

Results are contributed to the Intel-Arc-GPU-Benchmarks repo.

arcbench exports a Markdown file with:

• Repo-format table row
• Hardware/software manifest
• Exact command used
• Raw llama-bench output

Open a PR at the benchmarks repo and drop the exported file in the appropriate results/ subdirectory.

---

License

MIT