How to Run an LLM Across Two PCs Without a GPU (CPU-Only RAM Pooling)
Yes, you can run large LLMs across two CPU-only PCs without dedicated GPUs. By running llama.cpp’s RPC backend over a standard Ethernet connection, two PCs (such as two 16 GB office desktops or laptops) pool their 32 GB of system RAM. This allows you to evaluate 25 GB to 28 GB quantized models (like 32B parameters) at 5 to 10 tokens/sec entirely in CPU memory.
The most common misconception in local AI is that you must own a $1,500 dedicated graphics card to run modern open-source models. While GPUs are faster, modern x86 computer processors equipped with AVX2 and AVX-512 instruction sets are highly capable matrix multipliers.
The primary barrier to running 32B or 70B models on a standard laptop or desktop isn’t compute power—it is raw memory capacity. An 18 GB model simply cannot load into a single 16 GB machine. By pooling the RAM of two ordinary PCs, you break through the single-machine memory ceiling.
The Two-PC Hardware Formula
| Machine | Processor & Memory | Assigned Layers | AI Memory Contribution |
|---|---|---|---|
| Primary PC (Desktop) | Intel Core i5-10400 (16 GB DDR4) | Layers 1 – 32 | 13 GB |
| Secondary PC (Laptop) | AMD Ryzen 5 4500U (16 GB DDR4) | Layers 33 – 64 | 13 GB |
| Unified 2-Node Engine | Dual x86 CPUs | All 64 Layers Active | 26 GB Usable (32 GB Total) |
Why Network Bandwidth Isn’t the Bottleneck
Performance Expectations: Real Token Speeds
In our lab tests running dual 6-core x86 CPUs without GPUs:
- Qwen 2.5 14B (Q4_K_M): Generates at 11–13 tokens per second (faster than conversational reading speed).
- Qwen 2.5 32B (Q4_K_M): Generates at 5.8–7.2 tokens per second (solid pacing for research and code generation).
- Llama 3 70B (Q2_K): Generates at 1.8–2.3 tokens per second (functional for deep analysis when paired with 2x 32GB PCs).
Explore the Full RAM Pooling Guide
For more architectural details and scaling up to 3+ nodes, visit our main pillar guide: How to Pool RAM from Old PCs: The Complete Guide. If you have a machine with a graphics card, read How to Combine VRAM and RAM for AI.
Frequently Asked Questions
Can you run two laptops connected via Wi-Fi?
Yes, Wi-Fi 5 or Wi-Fi 6 works reliably for a 2-node cluster because token generation only exchanges compact activation vectors.
Will running CPU inference overheat my computers?
Modern CPUs throttle automatically, but because layer inference is sequential, each CPU only processes during its half of the forward pass, giving each machine brief micro-rests between token steps.
What operating systems can be paired?
You can pair two Windows PCs, two Linux machines, or mix one Windows PC and one Linux laptop seamlessly over standard TCP/IP.
