A GPU kernel is memory-bound when data arrives too slowly to keep its arithmetic units busy. Arithmetic intensity—the operations performed per byte moved—helps distinguish this regime from compute-bound execution. section 3.1.3 section 3.1.4
Arithmetic intensity relates a kernel’s computation to its data movement; lower values indicate less work per byte transferred. section 3.1.3
The roofline model plots arithmetic intensity against performance, with compute and memory ceilings. A kernel’s position indicates which resource limits it and how much optimization headroom remains. section 3.1.5
Common memory-bound workloads include:
- Element-wise operations, reductions, standalone softmax, and normalization, which perform relatively little arithmetic per byte moved. section 3.1.4
- The elementwise part of SwiGLU: intermediate writes and reads lower arithmetic intensity. A fused Triton kernel keeps activation and multiplication intermediates in registers, avoiding those global-memory transfers. section 3.3.1

















