Learn the scientific approach to GPU optimization: fix workload, inspect, classify limiter, change small, and measure.

Kernels, grids, blocks, and warps of 32 threads—and why divergence stalls the group.

How GPUs swap ready warps to hide memory wait times and why occupancy matters.

Efficient global-memory fetches when threads access contiguous addresses, and Structure of Arrays.

On-chip shared memory banks, conflicts when threads hit the same bank, and padding fixes.

Arithmetic intensity shows whether you are compute-bound or memory-bound on the roofline.

How GPU-resident fused MLPs cut memory traffic in a neural rendering pipeline.

Quality gates with FLIP and exceedance-rate validation for LLM-proposed optimizations.

Recap: performance loop, warps, memory ceilings, and evidence-based validation.