NVIDIA PAIR routes local AI requests across compatible computers

PAIR is software, not a physical router; it discovers and connects compatible computers on a network.[2] NVIDIA describes it as free, open-source software for putting compatible computers to work together on local AI.[1] It provides a local endpoint for AI applications and decides which connected computer serves each inference request.[3] PAIR works with Ollama and LM Studio by proxying their familiar interfaces, so agent applications do not need a new cluster API or changes to an agent harness.[4]

PAIR routes each independent request to an available system on the home network.[4] It assigns a request to one eligible computer, where the request runs from start to finish; it does not pool GPU memory or make several GPUs act like one larger GPU.[4][3] PAIR can adapt as devices join or leave the network.[1] A computer can serve a request only if it has the required inference engine enabled and the exact requested model available.[4] Adding computers can increase the number of requests running at once, but does not make an individual request faster.[3]

The beta is available for Windows, macOS and Linux through graphical and terminal interfaces.[4] Supported hardware includes GeForce RTX 20-series GPUs and newer, RTX PRO workstation GPUs based on Turing or newer, DGX Spark, and Apple M4-or-newer silicon.[4] The supplied requirements list Windows 11, DGX OS, Ubuntu 14.04 and macOS Tahoe as validated operating systems, and specify at least 8GB of RAM.[5] Internet access is not required to operate PAIR, but is needed to download models; PAIR can help install an inference engine and start model downloads on paired computers.[5][4] NVIDIA says special cables, racks and complex cluster infrastructure are not required.[5]

NVIDIA says PAIR is most useful for several independent requests, including multi-agent applications and concurrent local AI tools.[4] It can route work to other computers so a primary PC remains available for gaming, content creation or other interactive tasks.[4] NVIDIA says PAIR is designed to keep prompts, data and inference traffic on the local network, and uses mutual TLS encryption for communications between paired computers.[4] A three-device demonstration averaged 8 minutes 48 seconds, but NVIDIA cautioned that it was unofficial, configuration-specific and not a general benchmark or promise of linear scaling.[4]