GPU Architectural Evolution: From Graphics Processing to General-Purpose Parallel Computing

Abstract

Graphics processing units (GPUs) have evolved from specialized graphics accelerators into highly programmable parallel computing platforms in response to the increasing demands of real-time graphics and data-intensive computation. Early GPU architectures relied primarily on fixed-function graphics pipelines, whereas the introduction of programmable shader stages substantially improved computational flexibility. The subsequent development of general-purpose GPU (GPGPU) computing further extended GPU applications beyond graphics rendering to artificial intelligence (AI), machine learning, scientific computing, and large-scale data-centre workloads. This review examines the major stages of GPU architectural evolution, with particular attention to the transition from fixed-function pipelines to programmable shader architectures and subsequently to general-purpose parallel computing. Representative architectural developments are discussed in terms of programmability, parallel processing capability, memory organisation, and specialised computing units. The comparison indicates that GPU evolution has been characterised by increasing programmability, greater parallelism, and closer integration of workload-specific acceleration with general-purpose computing capabilities. Future GPU development is expected to place increasing emphasis on balancing computational throughput, memory efficiency, programmability, and specialised acceleration for increasingly diverse parallel workloads.