Unrolling thread...
From SIMT to Systolic: A Foundation for GPU and TPU Architecture