Vx achieves CUDA-like GPU kernel speed

The author compares GPU kernel performance in Vx against CUDA on an NVIDIA A100, finding that across-loops in Vx match or exceed CUDA speeds. Tests on a grid scaling loop and three LLM.c kernels (GELU, layernorm, and softmax) demonstrated competitive performance within 5% of optimized CUDA. Achieving this speed required lane-level reduction functions, fast math approximations, and a compiler optimization that embeds block sizes directly into generated GPU code. Open article →