Operator fusion merges consecutive computational operations in neural network graphs to reduce memory transfers between GPU global memory (HBM) and compute units, improving both speed and energy efficiency. Distinction from kernel fusion: operator fusion works at the computation graph level (merging graph nodes), while kernel fusion works at the GPU programming level (combining CUDA kernels). In practice, the terms are often used interchangeably. Fusion categories: (1) Element-wise fusion—combine sequential point-wise operations (add, multiply, activation) that share same tensor shape; (2) Reduction fusion—merge reduction operations (sum, mean, norm) with preceding element-wise ops; (3) Broadcast fusion—combine broadcast operations with subsequent computations; (4) Memory-intensive fusion—combine operations that are memory-bandwidth limited. Graph-level optimization: (1) Identify fusible operation sequences in computation graph; (2) Replace sequence with single fused node; (3) Generate optimized kernel for fused operation; (4) Eliminate intermediate tensor allocations. Framework implementations: (1) PyTorch Inductor (torch.compile)—automatic fusion with Triton code generation; (2) TensorRT—aggressive layer fusion for inference optimization; (3) XLA (JAX/TensorFlow)—HLO fusion passes; (4) ONNX Runtime—graph optimization including fusion; (5) Apache TVM—auto-tuned fused kernels. Performance impact by operation type: (1) Element-wise chains—2-5× speedup (dominated by memory); (2) Attention fusion—2-4× speedup and memory reduction; (3) Normalization + activation—1.5-2× speedup. Limitations: (1) Not all operations can be fused (data dependencies, different tensor shapes); (2) Complex fusion may reduce parallelism; (3) Custom kernels harder to debug and maintain. Operator fusion is a core optimization pass in every modern deep learning compiler and inference engine, essential for closing the gap between theoretical hardware performance and actual application throughput.