What kernel fusion removes

Synthetic English narration.

Three operations, one output
Consider a small elementwise calculation. Add two input arrays, multiply the result by two, then replace negative values with zero. The output is all we need.

Launch three kernels
A separate kernel performs each operation. Intermediate arrays are written to device memory, then read back by the next kernel. Each launch also has overhead.

Count the temporary traffic
The intermediate values carry useful information, but these full temporary arrays need not always exist. Here, two arrays connect three simple elementwise operations.

Fuse the operations
One fused kernel loads the inputs and carries intermediate values locally through all three operations. It writes the final output once, using one launch.

Check the answer and the cost
The output is unchanged in this integer example. Real floating-point kernels need numerical checks. Fusion can save launches and memory traffic, but its speed must be measured.