BROWSE
Detail
Conference Paper
A Bounded-Search of Size-Local Blind Spots in an Official cuTile GEMM Sample on Blackwell
Cited - time in 
Share 
Share
- Authors
- Jiwon Lee, JooHyoung Cha, Yongjoo Kim, Yongin Kwon
- Issue Date
- 2026-05
- Citation
- Annual Symposium of KIPS (ASK) 2026, pp.96-98
- Publisher
- 한국정보처리학회
- Language
- English
- Type
- Conference Paper
- Abstract
- We present a measurement study of FP16 square GEMM on an NVIDIA RTX PRO 6000 Blackwell GPU. Rather than identifying the universally fastest framework, we ask whether an official cuTile sample configuration can misrepresent framework capability at irregular sizes when stronger configurations exist in the same environment. We evaluate a bounded candidate pool of ten cuTile-Python configurations across five irregular sizes (1664, 1792, 1920, 2432, 2560). Using the official 128x256x64 regular configuration as the reference, measured regret ranges from 74.73% to 84.65% (mean 79.33%). At sizes 1664 and 2560, this gap is sufficient to reverse the relative cuTile-Triton ranking depending solely on which cuTile path is reported.