Independent evaluation of NVIDIA's CUDA Tile shows strong performance on Blackwell but reveals cross-architecture portability gaps compared to Triton

Researchers benchmarked CUDA Tile (CuTile), a Python-based abstraction for GPU kernel development, against cuBLAS, Triton, WMMA, and raw SIMT on three NVIDIA GPUs: H100 NVL, B200, and RTX PRO 6000 Blackwell Server Edition, testing GEMM (matrix multiplication), fused multi-head attention, and end-to-end LLM inference in BF16/FP16 precision.

On Blackwell B200, CuTile achieved up to 1007 TFLOP/s for fused attention, outperforming FlashAttention-2 by 2.5×, while GEMM reached 52-79% of cuBLAS performance in 22 lines of Python code (versus 123 for WMMA). On RTX PRO 6000, the same CuTile attention kernel achieved only 53% of FlashAttention-2 throughput, exposing significant cross-architecture optimization gaps.

Triton sustained 62-101% of cuBLAS performance across all tested platforms without architecture-specific tuning, demonstrating substantially stronger portability than CuTile.