GPU occupancy measures how fully an SM is populated with concurrent warps: active warps divided by the SM’s maximum supported warps. If an SM supports 64 warps but register usage permits 32, occupancy is 50%. Higher occupancy gives the scheduler more warps to switch between while others wait on memory, helping hide latency. section 3.2.1
In Triton, local variables are assigned to registers automatically. Increasing register usage—especially in fused kernels that keep more intermediates alive—can cross an occupancy threshold, reduce concurrent warps, and expose memory latency. Lower occupancy is not automatically worse: sufficient registers can prevent spills, while performance also depends on instruction-level parallelism, memory coalescing, and Tensor Core utilization. section 3.2.1 section 3.7
Register pressure in a fused kernel
Fusing operations can reduce global-memory traffic by keeping intermediate values in registers, but those additional live values may lower occupancy. The practical target is a balance: retain enough concurrent warps to hide latency without forcing register spills or starving each thread of working registers. section 3.2.1 section 3.7


