README.md

metadata

title: README
emoji: 🥞
colorFrom: pink
colorTo: green
sdk: static
pinned: false
short_description: Community LLM benchmarks on NVIDIA DGX Spark

⚡ Spark Arena

Real benchmarks. Real hardware. Real results.

Spark Arena is a community-driven LLM performance leaderboard for the NVIDIA DGX Spark (GB10). We benchmark large language models on real DGX Spark hardware across multiple inference runtimes — vLLM, SGLang, TensorRT-LLM, and llama.cpp — and publish transparent, reproducible results with full access to our recipes and configurations.

🏆 What we do

• Leaderboard — head-to-head LLM inference performance on DGX Spark: spark-arena.com/leaderboard
• Transparent methodology — every result ships with the recipe and configuration used to produce it
• Blog — deep dives on reproducibility and performance tuning: blog.spark-arena.com

🛠️ Open source

• spark-vllm-docker — Docker containers for LLM inference engines on DGX Spark
• llama-benchy — benchmarking tool for measuring inference performance
• sparkrun — CLI for managing workloads on DGX Spark systems

🤝 Community

Born in the NVIDIA GB10 / DGX Spark developer forums and maintained by the community.

• 💬 Discord
• 🐦 X / Twitter
• 🌐 spark-arena.com