MEMORY DEVICE FOR PERFORMING ASYMMETRIC COMPRESSION AND DECOMPRESSION AND OPERATING METHOD THEREOF

Smart overview of the Invention

The memory device described utilizes advanced neural network models to perform asymmetric compression and decompression on data values. It employs an invertible neural network model for generating latent data representations of data values intended for storage. Subsequently, a non-invertible neural network model is used to generate distribution parameters for these latent representations, which are then compressed accordingly. This approach aims to optimize data storage efficiency and retrieval processes within a memory space.

Technology Background

The invention operates within the context of high-speed interconnect technology, specifically leveraging Compute Express Link (CXL) for enhanced data transmission. CXL facilitates rapid communication between processors, memory, and accelerators, making it suitable for high-performance computing environments. Deep learning-based neural networks, with their ability to map nonlinear relationships in data, are employed for the compression tasks, showcasing recent advancements in data compression techniques.

Functional Components

The memory device comprises a memory array for storing compressed data, processors with processing circuitry, and a memory with stored instructions. The instructions enable the processors to perform domain conversion, reducing data entropy using an invertible neural network model. Distribution parameters are generated via a non-invertible neural network model, which then guides the compression process through entropy coding. This setup allows efficient data storage and retrieval based on host device commands.

Operation Method

The method involves generating latent data representations for data to be written in memory, using an invertible neural network model. The latent data is then compressed based on distribution parameters from a non-invertible model. Upon receiving a read request, the device partially decompresses the relevant latent data representation, regenerating the original data value for provision to the host device. This process ensures asymmetric compression and efficient data retrieval.

Additional Considerations

The detailed description emphasizes flexibility and adaptability in implementation, allowing for various modifications and alterations. Terminology used is clarified to ensure consistent understanding, and the invention is presented as adaptable to different embodiments. The description also outlines the structural and functional components, highlighting their potential configurations and interconnections, reinforcing the invention's applicability to diverse technological contexts.