Can Large Language Models Predict Parallel Code Performance?
Accurate determination of the performance of parallel GPU code typically requires execution-time profiling on target hardware - an increasingly prohibitive step due to limited access to high-end GPUs. This paper explores whether Large Language Models (LLMs) can offer an alternative approach for GPU performance prediction without relying on hardware. We frame the problem as a roofline classification task: given the source code of a GPU kernel and the hardware specifications of a target GPU, can an LLM predict whether the GPU kernel is compute-bound or bandwidth-bound?
Authors
- Kirk W. Cameron (ORCID: https://orcid.org/0000-0002-1679-5448)
- Gregory Bolet (ORCID: https://orcid.org/0000-0002-5009-0839)
- Gal Oren (ORCID: https://orcid.org/0000-0002-8831-5423)
- Harshitha Menon (ORCID: https://orcid.org/0000-0003-4707-9580)
- Konstantinos Parasyris (ORCID: https://orcid.org/0000-0002-8258-9693)
- Niranjan Hasabnis (ORCID: https://orcid.org/0009-0004-4010-7213)
- Giorgis Georgakoudis (ORCID: https://orcid.org/0000-0001-6542-3555)
- Hayden Estes (ORCID: https://orcid.org/0000-0002-3419-2632)
Institutions
- Lawrence Livermore National Laboratory (US)
- Metis Design Corporation (United States) (US)
- Virginia Tech (US)
- Stanford University (US)
Publication Details
- Published
- 2025-07-20
- DOI
- https://doi.org/10.1145/3731545.3743645
- Citations
- 1
- Primary Topic
- Parallel Computing and Optimization Techniques
- Type
- article
- Field-Weighted Citation Impact
- 2.27