cuDNN (CUDA Deep Neural Network library) is NVIDIA's GPU-accelerated library providing highly optimized, low-level implementations of the primitive operations that underpin deep learning — convolutions, pooling, normalization, activation functions, and recurrent operations — tuned specifically for NVIDIA GPU architectures. Deep learning frameworks including PyTorch, TensorFlow, and JAX call into cuDNN under the hood rather than implementing these operations themselves, which is why upgrading a GPU driver/cuDNN version can materially change training and inference performance without any change to model code. cuDNN's optimizations include algorithm auto-tuning (selecting the fastest available convolution algorithm for a given tensor shape and hardware), Tensor Core utilization for mixed-precision workloads, and kernel-level performance engineering that individual framework developers would find impractical to reimplement and maintain for every GPU generation.

The distractors point to different, specific NVIDIA-ecosystem or third-party tools: text-to-image generation via CLIP (A) is an application-level generative task, not a low-level compute library's function. GPU metrics monitoring via Prometheus (B) describes observability tooling (commonly paired with NVIDIA's DCGM exporter), a separate concern from computational optimization. GPU-accelerated data preparation (D) more closely describes RAPIDS libraries like cuDF, not cuDNN, which is specifically scoped to neural network primitive operations rather than general data preprocessing.