Connecting several spare PCs does not create one giant GPU. NVIDIA PAIR’s real job is to send independent AI requests to available PCs so they do not all queue up in front of a single GPU.
PAIR distributes requests; it does not combine GPUs
PAIR discovers computers on the same local network, checks each node’s inference engine, models, and workload, then routes requests to them. It is a local inference router. It currently supports Ollama and LM Studio, and provides applications with a single Ollama-compatible or OpenAI-compatible endpoint.
There is one crucial boundary here. PAIR does not pool VRAM or split one model across multiple computers. A single request is handled from start to finish on one selected node. So it is not for connecting two 24 GB GPUs to run a model that needs 48 GB. Instead, it is suited to distributing the queue of independent requests sent simultaneously by multiple agents or users.
| What you want to do | Is PAIR suitable? | Why |
|---|---|---|
| Handle requests from multiple sub-agents at once | Suitable | Places independent requests on different ready nodes |
| Distribute local inference requests from multiple apps | Suitable | Provides Ollama- and OpenAI-compatible proxy endpoints |
| Increase one request’s generation speed by the number of GPUs | Not suitable | Does not split an in-progress request across nodes |
| Pool VRAM across multiple GPUs to run a large model | Not suitable | Does not support VRAM pooling or model sharding |
In NVIDIA’s launch demo, five Hermes Desktop sub-agents used Qwen 3.6 35B A3B. They averaged 18 minutes on one RTX Spark laptop, versus 8 minutes 48 seconds across three nodes: that laptop, a DGX Spark, and an RTX 5090. NVIDIA also makes clear that this is an unofficial demo for a specific configuration. Results vary with parallelism, model, network, engine settings, and node state, so it should not be taken as a general 2× acceleration figure.
What determines performance is not the number of PCs, but replicas of the same model
Installing PAIR on several nodes alone does not automatically spread requests evenly. For a node to receive a request, it must be online, have a compatible engine running, and contain the exact model requested. If the same model tag is downloaded on only one of three nodes, requests for that model will ultimately converge on that one node.
Conversely, you do not need to put only identical models on every computer. You can keep a larger model on a high-memory workstation and a smaller model on a laptop, then direct requests by name. But if you want to increase concurrent throughput for a particular model, prepare the same model tag on every node assigned to that model.
For your first test, two similarly specced PCs are easier
The current scheduler mainly considers queued jobs and gently adjusted GPU utilization. It does not assess GPU model, available memory, whether the model is already loaded in memory, or how heavy a request will be. Rather than immediately mixing PCs with large performance gaps, start by confirming routing with two similar machines and a small shared model; it will be easier to identify the cause of issues.
You should also consider PAIR support separately from inference-engine support. PAIR itself runs on x64 and arm64 versions of Windows 11, Linux, and macOS, and can group nodes running different operating systems. NVIDIA’s listed validated hardware includes GeForce RTX 20 Series and later, RTX PRO from Turing onward, DGX Spark, and systems with Apple M4 or later. But whether a model can actually run depends on Ollama or LM Studio support for the operating system, GPU, and drivers, as well as enough memory to hold that model.
“Local” still requires a network you trust
Nodes find each other through mDNS, and you can add one by IP address if automatic discovery does not work. Entering the 6-digit PIN displayed by the inviting PC on the other PC establishes certificate trust, and mutual TLS is then used for key communication between nodes.
That does not mean it is safe to use on public Wi-Fi. The 6-digit PIN is a short code for the initial connection, not a long-term credential. NVIDIA’s security documentation says that some discovery and node-information traffic may be plaintext, and other devices on the same subnet may read hostnames, hardware details, and utilization. Pair only on a trusted home or private-office network, and do not expose PAIR endpoints through router port forwarding or an unauthenticated public reverse proxy.
Local inference is not the same as being completely offline
PAIR’s inference path is designed to remain on the local network, but downloading models requires internet access. Connected apps, model catalogs, update features, or separately configured engines may also contact external services. If you handle sensitive material, review not only PAIR but also client apps and model delivery paths.
A two-PC PAIR setup workflow
Check support and install it on each PC
Download the Windows executable, the .deb package for Debian-based Linux, or the macOS .dmg from the NVIDIA PAIR distribution page. On Linux, you can install it from your download folder with sudo apt install./NVPAIR-Setup-*.deb. RPM-based distributions currently require a source build. After installation, run PAIR on every computer and wait until the local node appears in Overview.
Start Ollama or LM Studio
Install an engine from the first-run window, or go to Overview → select node → Engine settings. If an existing installation of Ollama or LM Studio is running, PAIR may adopt it. If you installed one anew, check that its status has changed to running. PAIR is only a router and does not run models directly, so every worker node needs an engine.
Connect the second PC with a 6-digit PIN
Put both PCs on the same trusted local network. On the first PC, click Add node in the upper right, or open Settings → Cluster → Available nodes. Select the other PC and enter the 6-digit PIN shown on the first PC into the invitation window on the other PC. If automatic discovery is empty, add it by IP address and check that PAIR is allowed through the firewall. When finished, both devices should appear in Connected nodes.
Prepare the same test model on both nodes
Download the identical model tag through Engine settings → Add model on each node. At first, choose a small model that fits comfortably in both computers’ memory. If the engine requires an explicit load, run Load after downloading. Even a one-character difference in a model name may prevent it from becoming a candidate node for the same request.
Connect the endpoint and prove that distribution is happening
Copy the current address from Endpoints → API endpoints and place it in your client’s Base URL. By default, the Ollama proxy provides the Ollama API and OpenAI-compatible paths at http://127.0.0.1:11434, while the LM Studio proxy provides OpenAI-compatible paths at http://127.0.0.1:1234. If you changed the port, use the address shown on screen. Send several independent requests simultaneously, then check each job’s Ran on in Jobs. Distribution has truly occurred only if different nodes are recorded.
When validating, do not look only at tokens per second for a single request. Run the same batch of prompts on one node and on multiple nodes, then compare total completion time, wait time, failures, and output quality. Include situations where node state changes, such as a laptop sleeping or a game running. PAIR’s value shows up less in peak speed than in how much it reduces queues when multiple requests overlap.
If you want to dig deeper
Personal AI Router for Local Inference | NVIDIA PAIR — Check supported platforms, validated configurations, and official downloads. nvidia.com
NVIDIA PAIR Virtual Inference Router Expands Available Compute on Your Local Network — Explains the routing architecture, multi-agent demo, and boundaries of suitable workloads. developer.nvidia.com
Getting Started | NVIDIA Personal AI Router — Follow the actual menus from installation and PIN pairing to model preparation and API connection. docs.nvidia.com
Known issues — Check factors the current scheduler does not consider and platform-specific limitations before deployment. docs.nvidia.com
Security: NVIDIA/Personal-AI-Router — Explains PINs, mutual TLS, local endpoints, and LAN trust boundaries in detail. github.com
