The SIMT execution model compared with classic SIMD

A colleague argues that a CUDA warp is 'just SIMD with extra steps' since all 32 threads execute the same instruction at the same time on real hardware. What is the most accurate characterization of the actual distinction NVIDIA draws between SIMT and SIMD?

25 of 46 in GPU / CUDA