Soaring memory costs have forced hardware enthusiasts to look for creative alternatives to pricey upgrades. With DDR5 prices multiplying drastically over the past year, a small company called TradeMAV is crowdfunding a hardware hub designed to borrow memory from old PCs, Macs, Linux machines, and even Android phones. Listed at $109 for early backers on Indiegogo, the device called RAMDeck promises to bridge disparate hardware into a single memory pool capable of handling AI models that would otherwise crash a single system.
Despite the ambitious marketing terminology, the underlying technology relies on an existing, free open-source framework. The RAMDeck software agent initiates the llama.cpp remote procedure call backend on each contributing device, splitting model layers across different machines. While the hardware hub handles automatic device discovery, layer placement prioritizing fast GPUs, and an OpenAI-compatible API, its internal hardware specifications remain largely undisclosed. TradeMAV's public documentation notes that users comfortable with terminal scripts can achieve similar functionality using the free open-source route.
Performance across a distributed network of devices inherently involves major trade-offs. TradeMAV's benchmarks reveal that while lighter mixture-of-experts models can maintain decent token generation speeds across multiple systems, denser models slow down drastically. A 27-billion dense model managed roughly two tokens per second across a four-device cluster, and a 120-billion parameter model crawled at just over one token per second after a lengthy initialization period. For advanced users, free alternatives like llama.cpp and similar open-source clustering tools offer a zero-cost path to experimentation, while the RAMDeck project primarily sells turnkey convenience alongside the standard risks associated with crowdfunding campaigns.
Crowdfunded RAMDeck Hub Pools Spare Device Memory to Run Larger AI Models
Comments