Abstract
GPU kernels generated by large language model (LLM) agents can remain less efficient than expert implementations, but runtime alone does not reveal how the gap relates to design discovery and implementation. We introduce D2K-Bench, a diagnostic benchmark of 26 tasks and 85 workloads that measures how effectively agents translate expert design guidance into efficient GPU kernels. The guidance covers L1: high-level algorithmic insights, L2: dataflow design, and L3: low-level optimization tricks, including dependencies among these levels. Pairwise runs with and without guidance share task descriptions, workloads, tools, hardware, and a 350-turn budget. Complementary assessments examine independently proposed designs and the design properties implemented in generated code. Across five models on NVIDIA B200 GPUs, guidance raises correctness over 130 model-task pairs from 93.1% to 98.5% and increases the Performance Score over all 26 tasks from 1.46 to 1.95. For the three frontier models with correct submissions on all 26 tasks in both runs (GPT-6-Astra, Claude-Opus-4.8, and GPT-5.6-Sol), geometric mean speedup increases from $1.69\times$ to $2.49\times$. Across all five models, the mean combined implementation score increases from 57 to 70 out of 100. These results show the value of expert design guidance while identifying design properties that remain unimplemented.
Keywords
Subject
Publication details
- Journal
- Not available
- Open access
- Green open access
Cite this article
APA 7
Li, D., Jiang, H., Zhang, C., Wu, W., Guo, X., Tu, J., Zhang, J., Yuan, B., & Liu, D. (2026). D2K-Bench: Can LLM Agents Turn Expert Designs into Efficient GPU Kernels? https://omanscience.com/en/articles/d2k-bench-can-llm-agents-turn-expert-designs-into-efficient-gpu-kernels
MLA 9
Li, Daifeng, et al. "D2K-Bench: Can LLM Agents Turn Expert Designs into Efficient GPU Kernels?" https://omanscience.com/en/articles/d2k-bench-can-llm-agents-turn-expert-designs-into-efficient-gpu-kernels.
Chicago (author–date)
Li, Daifeng, Huiqiang Jiang, Chengruidong Zhang, Wei Wu, Xudong Guo, Jianhong Tu, Jianwei Zhang, Binhang Yuan, and Dayiheng Liu. 2026. "D2K-Bench: Can LLM Agents Turn Expert Designs into Efficient GPU Kernels?" https://omanscience.com/en/articles/d2k-bench-can-llm-agents-turn-expert-designs-into-efficient-gpu-kernels.
Harvard
Li, D., Jiang, H., Zhang, C., Wu, W., Guo, X., Tu, J., Zhang, J., Yuan, B. and Liu, D. (2026) 'D2K-Bench: Can LLM Agents Turn Expert Designs into Efficient GPU Kernels?', Available at: https://omanscience.com/en/articles/d2k-bench-can-llm-agents-turn-expert-designs-into-efficient-gpu-kernels.
Vancouver
Li D, Jiang H, Zhang C, Wu W, Guo X, Tu J, et al. D2K-Bench: Can LLM Agents Turn Expert Designs into Efficient GPU Kernels? https://omanscience.com/en/articles/d2k-bench-can-llm-agents-turn-expert-designs-into-efficient-gpu-kernels
IEEE
D. Li, H. Jiang, C. Zhang, W. Wu, X. Guo, J. Tu, J. Zhang, B. Yuan, and D. Liu, "D2K-Bench: Can LLM Agents Turn Expert Designs into Efficient GPU Kernels?," https://omanscience.com/en/articles/d2k-bench-can-llm-agents-turn-expert-designs-into-efficient-gpu-kernels.