Nvidia has introduced PAIR, a utility designed to distribute AI agent tasks across multiple GPUs within a local network. By utilizing idle compute cycles from various machines, the tool aims to prevent performance bottlenecks that occur when multiple sub-tasks are processed on a single graphics card.
The system operates by creating a proxy for AI front-ends like Ollama and LM Studio. It orchestrates sub-tasks across available nodes, which must also have PAIR and the relevant AI software installed. The tool automatically handles model downloads and uses mDNS or IP address discovery to connect systems.
Designed for elasticity, PAIR adapts to changing resource availability. If a user reclaims their GPU for gaming or other tasks, the system adjusts its workload distribution accordingly. While it does not guarantee specific quality of service, it is intended for long-running tasks that do not require strict deadlines.
The software supports a wide range of hardware, including Nvidia GeForce RTX 20-series cards and newer, as well as DGX Spark systems. It also provides support for Macs equipped with M4-series processors. The PAIR client is compatible with Windows, macOS, and Linux operating systems.