Open benchmark leaderboard for AI accelerators on LLM workloads.

→ Live Leaderboard · Contributing · Suites · Discussions · Development

From workload spec to published result — every row on the leaderboard carries its runner hash, environment fingerprint, and accuracy receipt.

---

Why AccelMark?

	The problem	AccelMark's answer
MLPerf	Rigorous but slow — only large vendors participate	Community runs often finish quickly (e.g. Suite A default ~11 min; Suite D default ~22 min; full all-scenarios run ~7 h)
Vendor whitepapers	Different setups make cross-vendor comparison impossible	Fixed schema + shared LoadGen = apples-to-apples
Most benchmarks	Cover only NVIDIA and only throughput	NVIDIA, AMD, Huawei Ascend, Apple Silicon — throughput, latency, scaling, quantization

---

Quick start

# 1. Clone and install
git clone https://github.com/FreedomIntelligence/AccelMark.git
cd AccelMark
pip install "jsonschema>=4.0""numpy>=1.24""pyyaml>=6.0"# framework dependencies (Python >=3.10 required)
pip install -r runners/nvidia_vllm_47f5d58e/requirements.txt # installs runner dependencies# 2. One-time setup
cp configs/submitter.yaml.example configs/submitter.yaml
# Edit configs/submitter.yaml — add your name# 3. Run the benchmark (~11 min on A100)
python run.py --runner nvidia_vllm_47f5d58e --suite suite_A
# 4. Submit your result — open a pull request:# git checkout -b submit/<your-hardware># git add results/community/<run_name>/ && git commit -m "results: <hardware>"# gh pr create # or open via the GitHub web UI## <run_name> is the directory auto-created by run.py — it already contains# your result.json and env_info.json; no manual file moves are needed.

See CONTRIBUTING.md for the full guide.

---

Suites

Suite	Model	Chips	Question answered	Primary metric
A	Llama-3-8B	1	How fast is this chip at inference?	Offline tokens/sec
B	Llama-3-70B	flexible	Can this chip serve large models?	Offline tokens/sec
C	Llama-3.1-8B	1	Quantization speed/quality tradeoff?	Speedup vs BF16
D	Llama-3.1-8B	1	How does this chip handle long-context (28K) inputs?	Offline tokens/sec
E	Llama-3-8B	1×/2×/4×/8×	How well does this chip scale?	Scaling efficiency
F	Qwen2.5-0.5B	1	How fast is this consumer/edge GPU?	Offline tokens/sec
G	Mixtral-8x7B-Instruct	≥2 (auto)	How efficiently does this chip handle sparse MoE inference?	Offline tokens/sec

Suites A, B, and D also include optional speculative decoding and/or burst load extra scenarios — see suites/README.md for per-suite details.

See suites/README.md for full specs, time budgets, SLA definitions, and metric descriptions.

---

Currently on the leaderboard

A snapshot of accelerators that have at least one submission on the leaderboard. Tile size is proportional to submission count; colour denotes vendor. See the live leaderboard for current rankings, per-suite breakdowns, and the underlying result.json files.

---

Supported platforms

Reference runners live under runners/ (see each folder’s meta.json). The table below is auto-generated from each runner's meta.json — never hand-edited. Add a runner, declare its suite_support in meta.json, and the matrix updates on its own.

Hardware	Runner folder	Framework	A	B	C	D	E	F	G	H
NVIDIA A100-SXM4-80GB	nvidia_sglang_inference_opt_5fe06410	SGLang	⋯	⋯	⋯	⋯	⋯	✓	⋯	?
NVIDIA GPU	nvidia_sglang_c43a8309	SGLang	✓	✓	✓	✓	✓	✓	✓	⋯
NVIDIA GPU	nvidia_tensorrt_llm_21323970	TensorRT-LLM	✓	⋯	✓	—	⋯	✓	⋯	⋯
NVIDIA GPU	nvidia_vllm020_0f6c56e4	vLLM	⋯	⋯	⋯	⋯	⋯	⋯	⋯	?
NVIDIA GPU	nvidia_vllm_47f5d58e	vLLM	✓	✓	✓	✓	✓	✓	✓	✓
NVIDIA V100 (SM70)	nvidia_onecat_vllm_12a253c2	1Cat-vLLM	⋯	⋯	⋯	⋯	⋯	—	⋯	?
AMD GPU	amd_vllm_rocm_6c18cd8f	vLLM-ROCm	✓	✓	✓	✓	✓	✓	✓	?
Huawei Ascend NPU	ascend_vllm_ascend_b1957851	vllm-ascend	✓	✓	✓	✓	✓	—	—	?
Huawei Ascend NPU	ascend_vllm_ascend_d4aa9fda	vllm-ascend	✓	✓	✓	✓	✓	—	—	?
Apple Silicon	apple_mlx_lm_9546b8b5	mlx-lm	⋯	—	—	⋯	—	⋯	—	?
Google TPU	google_vllm_tpu_68cc9ffa	vllm-tpu	✓	—	—	✓	—	✓	—	?
Moore Threads GPU	moorethreads_vllm_musa_f2f6f965	vllm-musa	✓	⋯	⋯	⋯	⋯	✓	—	?

Legend: ✓ validated · ⋯ author-declared (not smoke-tested in this repo yet) · — unsupported.

Regenerate locally with python tools/generate_platforms_matrix.py. CI runs --check and fails the PR if the README and runner metadata disagree.

Other stacks (TensorRT-LLM, MindIE, mlx-lm, etc.) can be added as new runner folders; see the contributor guide.

Adding a new runner? See CONTRIBUTING.md#adding-a-new-runner. Adding a new accelerator family? See runners/README.md.

---

Leaderboard tiers

Tier	How	Where
community	Submitted by anyone via PR and passes CI validation	Community tab
verified	Independently reproduced on the same hardware/runner and matches the original within 5%	Main leaderboard

Community results are fully visible and comparable — they just haven't been independently reproduced yet. Anyone with the listed hardware can promote a community result to verified by submitting a reproduction PR.

---

Contributing

The most valuable contribution is running the benchmark on hardware not yet in the leaderboard.

• Submit a result → Submitting a result
• Add a new runner → Adding a new runner
• Add a new accelerator family → Platform plug-in guide
• Report a bug → Open an issue
• Ask a question / share results → Discussions
• Extend the leaderboard → Development guide

Optional: AccelMark also ships a small voice-driven launcher for the OpenClaw ecosystem — see openclaw_skill/. It's not required to run, contribute, or submit results.

---

Citation

If you use AccelMark results in research, please cite:

@misc{accelmark2026,
title = {Beyond NVIDIA! A Multi-Regime Framework for Benchmarking Heterogeneous AI Accelerators},
author = {Liang, Juhao and Zhang, Zhiyuan and Li, Siyu and Lin, Zhihang and Yu, Minchen and Zeng, Li and Chen, Zizhong and Sun, Ruoyu and Wang, Benyou},
year = {2026},
url = {https://github.com/FreedomIntelligence/AccelMark}
}

---

License

Apache 2.0 — see LICENSE. Submitted benchmark results are contributed under CC BY 4.0. Bundled third-party data (datasets, accuracy subsets) keeps its upstream license — see NOTICE.