One Cluster, Every Setup: Why RAMDeck Fits Wherever You Start
Whether you have a spare laptop, a stack of Macs, a gaming GPU or a fast NVMe, RAMDeck turns what you already own into one AI cluster that runs models no single machine can hold.
Models keep getting bigger. The memory in one box doesn’t. RAMDeck closes that gap by pooling the RAM, VRAM and compute of every Windows PC, Mac, GPU box and phone on your network into a single cluster. Plug in the hub, install a node on each machine, and a 27B, 40B or 120B model runs across devices that could never load it alone.
Here is how RAMDeck works with every setup people are asking about.
Start With a Drawer of Old Computers
This is where RAMDeck began. A 2019 12GB Acer laptop with a spinning hard drive served as the primary node and ran a 13B model at 11-14 tok/s, reaching 19 after sharding improvements. With the Mac mini and the RTX 3060 mini PC behind it, the same laptop later ran a dense 27B at 1.92 tok/s, a model whose roughly 16GB it could never hold alone. Every byte sat in pooled RAM, with no disk in the generation loop. Put a stronger machine in the primary seat and speeds climb: a 40B MoE model ran at 16 tok/s with the mini PC as primary.
Already Own a Strong Machine? Add to It
A good GPU box gets you far, until the next model doesn’t fit. When that day comes, RAMDeck doesn’t ask you to replace it. Add the laptop from the closet, and the combined memory takes over. Our RTX 3060 mini PC ran a 7B coding model at 43-44 tok/s with a Mac coding against it from across the room, then led the cluster as primary for a 25GB 40B model.
GPU Owners: Pool Cheap Cards Instead of Buying One Giant One
GPU memory has become expensive. The RTX 5090 launched at $1,999 with 32GB, and street prices now run roughly $5,000-$7,200. Used RTX 3060 12GB cards average about $205. Three of them give you 36GB for a fraction of the price per gigabyte. With RAMDeck, each card lives in its own PC, and the hub makes them one pool. Add a card and your ceiling rises.
Mac Owners: One Cluster, Any Operating System
Mac-first tools shine when every device is a Thunderbolt 5 Mac. exo’s fast path uses RDMA over Thunderbolt 5, which needs TB5 Macs and compatible cables. RAMDeck takes the machines you actually have: an M4 Mac mini, an M3 MacBook, a 2017 Intel MacBook, a Windows gaming PC, an Android phone, over ordinary Ethernet or Wi-Fi. Our 120B run used exactly that mix.
Running MoE Models? RAMDeck Does Those Too
Mixture-of-experts models are efficient per token, but you still need somewhere to hold them. RAMDeck ran a 40B MoE at 16 tok/s and gpt-oss 120B, with 117B parameters and 5.1B active per token, across six devices at 1.37 tok/s. Five devices held layers, and the sixth, a 2017 Intel MacBook, joined the cluster and simply wasn’t needed. RAMDeck uses what the model requires and no more.
Dense Models Are Where Pooled RAM Shines
A dense model activates every parameter on every token, so it has no small set of weights to cache. Our dense Qwen3.8 27B ran with the entire model resident across four devices, led by the weakest laptop we own. Pooled RAM keeps every weight a memory-access away, which is why dense models are a natural fit.
Already Using llama.cpp RPC? RAMDeck Is the Easy Button
The layer splitting comes from llama.cpp’s RPC backend, and the repo documents the stock llama-cli –rpc path so you can verify it yourself. RAMDeck adds everything around it:
- Discovery of your devices on the LAN
- A model-fit check before you load
- Per-node memory targets
- Rebalancing when a device joins or drops, with automatic recovery
- A dashboard, an OpenAI-compatible API and a one-click benchmark
- A knowledge base sharded across the cluster’s RAM
The engine is source-available, so you can read it before you buy anything.
Wherever You Start, Capacity Scales With What You Plug In
| Your starting point | What RAMDeck adds |
|---|---|
| A few spare laptops and PCs | Models that none of them can load alone |
| One strong GPU box | More capacity the moment a model outgrows it |
| Several GPUs in several PCs | One pool of VRAM that grows card by card |
| A mix of Macs and Windows | One cluster across operating systems |
| A shelf of mismatched hardware plus a phone | A working cluster on day one |
Speed scales with what you plug in. Wire the primary, put a GPU in the pool, and tokens per second climb. Every cluster ships with a one-click benchmark, so you measure your own numbers instead of trusting ours.
RAMDeck is live on Indiegogo: https://www.indiegogo.com/en/projects/trademav/ramdeck-pool-your-ram-run-bigger-ai
Engine source: https://github.com/trademav/ramdeck-core-public- The TradeMAV Team