Evaluating CUDA Tile for AI Workloads on Hopper and Blackwell GPUs

Back to Feed

The paper evaluates NVIDIA's CUDA Tile (CuTile), a Python-based, tile-centric GPU programming abstraction, across Hopper and Blackwell GPUs and compares it with cuBLAS, Triton, WMMA, and raw SIMT. Benchmarks cover GEMM, fused multi-head attention, and end-to-end LLM inference in BF16/FP16 on H100 NVL, B200, and RTX PRO 6000 Blackwell Server Edition, finding CuTile's performance is highly workload- and architecture-dependent, with Blackwell B200 achieving up to 1007 TFLOP/s for fused attention (2.5x FlashAttention-2) using about 60 lines of Python. GEMM runs at 52-79% of cuBLAS with 22 lines of CuTile code versus 123 for WMMA; on RTX PRO 6000 the CuTile attention kernel reaches only 53% of FlashAttention-2, while Triton sustains 62-101% of cuBLAS across platforms, indicating better portability.

Discussion (0)

No comments yet