TiledAttention on NVIDIA DGX GB10: Supplementary Benchmark and Nsight Compute Results
收藏资源简介：
Description: This Zenodo record provides supplementary materials for the TiledAttention study, containing the full contents of benchmark-gb10/results from the TiledAttention repository. It includes benchmark and tile-tuning outputs (benchmark_results.csv, tuning_results.csv), reproducibility and sensitivity tables (table3_reproducibility.md, table4_tiling_sensitivity.md), study summary metadata, and Nsight Compute profiling artifacts (.ncu-rep and raw CSV exports) for representative attention workloads. The benchmarks cover sequence lengths $S={512,1024,2048,4096,8192}$, head dimensions $D={64,96,128,160}$, $float16$/$bfloat16$, and causal/non-causal modes; the baseline used in this environment is PyTorch SDPA (torch_sdpa). These files are intended to support reproducibility, independent analysis, and figure/table verification for the associated manuscript. Code: https://github.com/thisistaimur/TiledAttention/
