Direct Answer: You cannot physically merge RAM across different motherboards into a single OS memory bank. But for running local AI, you can pool memory over your local network: software splits an AI model’s layers across computers on your LAN, so each machine’s RAM holds a slice of the model. Three 16 GB computers can pool 42 GB of usable memory to run 32B to 70B models that none could load alone.
Why Forum Answers Say “No” (and Why They Are Right for Gaming)
When PC enthusiasts search forums, every post states that sharing RAM across separate motherboards is impossible. For traditional desktop applications and gaming, that answer is 100% correct. CPU memory controllers communicate across copper motherboard traces at speeds between 25,000 and 50,000 MB/s with sub-100 nanosecond latency. Standard Gigabit Ethernet runs at 125 MB/s with 0.5 to 1 millisecond latency—roughly 10,000 times too slow for a video game kernel.
Why Local AI Inference Works Over Ethernet
Large Language Models (LLMs) calculate sequentially through transformer layers. Weights only load into each computer’s RAM once at startup over TCP. During active token generation, each machine evaluates its assigned layers and passes only a compact intermediate activation vector (approximately 650 KB per token) over the network to the next machine.
Because only tiny activation vectors cross the network during token generation, gigabit Ethernet is more than fast enough. The primary bottleneck is network latency (ping) and per-device compute, not raw bandwidth.
Real-World Tested Benchmarks: The 120B Model Cluster
In verified benchmark tests, a heterogeneous cluster of 6 mismatched consumer devices (an RTX 3060 mini PC, a 12GB Acer laptop, an M4 Mac mini, an M3 MacBook, a 2017 Intel MacBook, and a Samsung Galaxy S24+ phone) successfully pooled memory to run a 60 GB 120B model (gpt-oss-120b) continuously for 40 hours over Wi-Fi and Ethernet at 1.37 tokens per second.
The system demonstrated automatic fault recovery: when an Android phone dropped offline and a laptop lid was closed mid-run, the cluster automatically redistributed remaining layers without crashing the session.
Read the comprehensive engineering guide, memory allocation math, and DIY open-source setup at RAMDeck.com.
Leave a Reply