CUDA kernels require time-consuming operations such as JIT compilation and kernel selection upon first execution, causing abnormally high inference latency for the first batch, which can be mitigated by using dummy inputs to trigger initialization in advance.