multiGPU

MultiGPU refers to using two or more graphics processing units in a single computer to increase graphics throughput or compute performance. It is used in gaming rigs, professional workstations, and data centers for tasks such as visualization, scientific simulations, and machine learning. The effectiveness of a multi-GPU setup depends on the workload and on software able to distribute work across GPUs.

Two common approaches exist. In consumer graphics, multi-GPU has historically relied on driver frameworks (for example, NVIDIA SLI and AMD CrossFire) to coordinate rendering across GPUs. More scalable interconnects, such as NVLink, are used in professional and data-center GPUs to improve bandwidth and coordination. Outside rendering, GPUs can cooperate on non-graphics compute tasks via GPGPU, splitting the workload or sharing data.

Software support is crucial. Graphics APIs like DirectX and Vulkan include multi-GPU features, but adoption varies by game and driver. For computation, frameworks such as CUDA and ROCm, along with machine-learning libraries (TensorFlow, PyTorch), implement data-parallel or model-parallel strategies across GPUs, with synchronization and memory sharing managed by the runtime.

Performance and scaling vary widely. Some workloads scale well with fast interconnects and well-optimized software, while others experience diminishing returns due to bandwidth limits, software limitations, or overhead from synchronization and frame pacing in games. Power, cooling, and driver stability are additional considerations.

Current trends emphasize specialized multi-GPU configurations in servers and professional workstations, while consumer multi-GPU gaming has become less common due to limited software support and efficiency concerns.