NVIDIA Research

# AI Research Benchmarks

Explore benchmarks from NVIDIA Research. Find evaluation tasks, understand scoring methods, and access results and submission guidance.

 Available benchmarks

## What can you evaluate?

Choose a benchmark to explore its tasks, methodology, and results.

NVIDIA Research

### SOL-ExecBench

Evaluate GPU kernels against theoretical hardware performance limits. SOL-ExecBench runs submissions on NVIDIA B200 GPUs and reports a Speed-of-Light (SOL) Score to help assess optimization headroom.

[Explore benchmark](https://research.nvidia.com/benchmarks/sol-execbench)

[View on GitHub](https://github.com/NVIDIA/SOL-ExecBench)

Available

  Evaluates GPU kernel performance

 Hardware NVIDIA B200

 Access NVIDIA account required to submit

 [Read the methodology](https://research.nvidia.com/benchmarks/sol-execbench/blog/introducing-sol-execbench)

NVIDIA Research

### SWE-Serve

Evaluate coding agents on 53 SGLang tasks covering model integration, serving APIs, cache systems, and GPU kernels. Compare task success with inference cost, agent steps, and wall time. Hidden tests check behavior, including live model serving.

[Explore benchmark](https://research.nvidia.com/benchmarks/swe-serve)

[View on GitHub](https://github.com/NVIDIA/swe-serve)

Available

  Evaluates Coding agents for LLM inference engineering

 Tasks 53 SGLang tasks · 6 engineering areas

 Hardware 12 CPU tasks · 41 single-H100 tasks

 Common questions

## Plan your evaluation

Check task coverage, access requirements, and scoring rules before comparing results.

 ### Which benchmark fits my evaluation?

 + Use SOL-ExecBench to evaluate GPU kernel performance on NVIDIA B200 hardware. Use SWE-Serve to evaluate coding agents on SGLang inference engineering tasks. Check that the tasks, hardware, and scoring method match your research question.

 ### Where can I find tasks and scoring details?

 + Open SOL-ExecBench to browse its Problems and Leaderboard sections. Its introduction explains the SOL Score. SWE-Serve includes a leaderboard, efficiency comparisons, and an explanation of task construction and closed-book evaluation. ```

 ### Do I need an account to submit results?

 + SOL-ExecBench requires an NVIDIA account to submit solutions and track your progress. You can browse its public benchmark information and leaderboards without signing in.

 ### How should I compare benchmark results?

 + Compare results under the same tasks, hardware, evaluation version, and scoring rules. Read the benchmark methodology before interpreting a score; results from different benchmarks are not directly interchangeable.