GPUonly

GPUonly is a term used to describe computer systems, software architectures, or workflows where graphics processing units (GPUs) provide the primary compute capability, with the CPU acting mainly as an orchestrator and host for I/O rather than performing central computations. In a GPUonly approach, kernels written for execution on GPUs implement the critical data-processing tasks, while the CPU handles scheduling, data preparation, and coordination between devices.

The concept has roots in the rise of general-purpose GPU computing (GPGPU) in the late 2000s and has become common in high-performance computing and machine learning workflows. While GPUs can deliver very high throughput for parallelizable workloads, true GPUonly operation is rarely practical for full systems because CPU-side control logic, memory management, and system I/O are often needed. Many deployments are effectively GPU-dominated, with the CPU performing lightweight control and data transfer.

Implementation relies on GPU programming models such as CUDA or OpenCL, and increasingly on domain-specific libraries and compilers that emit GPU kernels. Modern GPUs offer large memory bandwidth and specialized units (tensor cores, ray-tracing units) that can accelerate a range of tasks, from matrix operations to graphics rendering. Limitations include data transfer bottlenecks, debugging complexity, and memory constraints.

In practice, GPUonly informs architectural decisions when throughput is paramount, but engineers typically maintain some level of CPU involvement for orchestration and system management. See also GPU computing, GPGPU, CUDA, OpenCL.