Computational Object Storage for ML Data Preprocessing Offload

The GPU is made multi-functional by enabling it to perform both traditional graphics processing and data preprocessing operations. The system configures the GPU to execute preprocessing tasks (such as image normalization, resizing, and augmentation) that were traditionally performed by dedicated preprocessing accelerators, thereby eliminating the need for separate preprocessing hardware while maintaining preprocessing capability.

The patent merges the preprocessing function with the GPU processing function. Instead of having separate preprocessing accelerators and GPUs, the system combines these functions into a single GPU unit, allowing the same hardware resource to serve dual purposes: preprocessing data before it reaches the training pipeline and performing graphics/compute operations when needed.

The system performs preprocessing operations in advance, before data enters the main training pipeline. By pre-processing data (normalization, augmentation, etc.) before it reaches the GPU for training operations, the system prepares data upfront so that the GPU can focus solely on high-performance training computations without being burdened by preprocessing tasks during critical training operations.

The GPU's function is made dynamic and adjustable. The system can configure the GPU to perform preprocessing operations when needed, and switch to traditional graphics/compute operations at other times. This dynamic allocation allows the GPU to adapt its role based on system demands, maintaining processing bandwidth for training while enabling preprocessing capability when required.