NVIDIA PAIR turns every device on your LAN into an AI co-processor

BERLIN — At the IFA technology trade show in Berlin, NVIDIA announced NVIDIA PAIR, an open-source virtual inference routing software designed to eliminate local hardware bottlenecks by distributing AI agent workloads across connected devices on a local area network. The emergence of agentic AI frameworks—where autonomous subagents perform specialized tasks simultaneously—has created severe computational bottlenecks for local setups. Rather than serving as an independent artificial intelligence model, NVIDIA PAIR acts as a lightweight proxy backend.

Running multiple subagents on a single workstation frequently saturates graphics processing unit (GPU) memory and processing capacity, causing severe system slowdowns. As multi-agent artificial intelligence systems increasingly overwhelm individual desktop hardware, technical infrastructure is shifting toward localized distributed computing to handle concurrent model execution. It sits between user-facing inference applications and the underlying hardware across a local network.

The system functions transparently behind existing graphical user interfaces and popular local artificial intelligence environments, such as Ollama and LM Studio. When deployed, PAIR scans the local area network to identify connected devices with available compute capability. Because the software operates as a network routing layer, developers and end users do not need to modify their subagents, alter model execution code, or adjust established software workflows.

Upon establishing secure network connections, it transforms these external machines into active compute nodes. When an agent generates multiple concurrent tasks or subagent queries, PAIR distributes these requests across the available network nodes instead of directing all tasks to a single host GPU. This distributed routing mechanism reduces task queues and lowers execution latency while freeing system resources on the primary machine.

NVIDIA has released PAIR as an open-source project hosted on GitHub, currently categorized in a public beta phase. To establish a distributed local compute cluster, users must install the software on both the primary host machine and all secondary network systems intended to serve as processing nodes. As a result, users can run intensive foreground software—such as high-end video games, 3D rendering applications, or compilation tools—on their main computer while secondary network systems process background AI workloads.

NVIDIA PAIR supports a wide range of hardware ecosystems across multiple operating systems, including Microsoft Windows, Linux distributions, and Apple Mac computers powered by Apple Silicon M4 chips or higher. On the graphics hardware side, compatibility extends backward to NVIDIA’s GeForce RTX 20 series GPUs. The RTX 20 series, originally released in late 2018 based on NVIDIA’s Turing architecture, marked the company’s hardware transition toward AI-accelerated computing through dedicated hardware Tensor Cores.