As soaring global memory prices squeeze home AI enthusiasts, startups are searching for hardware workarounds to run larger models locally. TradeMAV has launched an Indiegogo crowdfunding campaign for its RAMDeck hub, priced at $109 for early backers, according to TechRadar coverage. The project aims to combat skyrocketing DDR5 prices—which industry reports trace back to intense AI datacenter demands for high-bandwidth memory—by linking old Windows PCs, Macs, and Android devices into a unified local memory pool. GitHub disclosures from the creators reveal that the team originally built the system out of personal necessity to run local assistant tools after outgrowing a single Mac Mini during the ongoing memory shortage.
Under the Hood and Technical Architecture
While marketed as a hardware memory-pooling solution, technical documentation shows that RAMDeck relies heavily on existing open-source plumbing. TechRadar and GitHub repository disclosures indicate that the engine underneath utilizes llama.cpp's ggml-rpc-server backend. TradeMAV's public source-available release exposes the core node-agent software, which runs on specific ports to shard large language model inference across heterogeneous hardware. Rather than performing low-level hardware merging, the system splits model layers across connected devices, passing intermediate results over the home network during token generation. TradeMAV’s proprietary contribution lies in orchestration wrappers, including automatic node discovery, intelligent layer placement favoring fast GPUs over system RAM, failure recovery, a dashboard, and an OpenAI-compatible API layer.
Performance Trade-Offs and Platform Constraints
Running large models across disparate consumer hardware introduces distinct performance penalties. Benchmarks cited by TechRadar indicate that while mixture-of-experts models paired with dedicated graphics cards can achieve reasonable speeds—such as 39 tokens per second for a 30B model distributed across a mini PC and an M4 Mac mini—dense models experience severe bottlenecks. A 27-gigabyte dense model dropped to under two tokens per second across four devices, and a larger parameter model crawled at similar rates after a lengthy initialization period. Furthermore, platform support comes with caveats. While the product aims for broad compatibility, the reference code released under an Apache 2.0 license carries a Commons Clause restriction against commercial resale, and current node agents lack native Linux support while relying on POSIX process APIs that limit Windows diagnostic scripting.
Evaluating the Trade-Off Between Hardware Convenience and Open-Source DIY
The emergence of RAMDeck highlights a broader tension within the local AI community between paying for convenience and leveraging freely available open-source tools. With consumer DRAM costs multiplying significantly since mid-2025 and projections indicating memory scarcity could persist for years, developers and enthusiasts are naturally looking beyond traditional hardware upgrades. However, as TradeMAV’s own documentation concedes, users comfortable with terminal scripts can achieve similar distributed inference setups entirely for free using stock llama.cpp RPC commands or alternative orchestration frameworks like exo. The $109 crowdfunded price tag primarily buys packaged orchestration, an API layer, and user-facing dashboards rather than novel AI or memory virtualization technology. Potential backers must weigh whether a pre-1.0 hardware wrapper justifies the financial outlay and inherent crowdfunding risks—such as delivery delays or unfulfilled specifications—when the underlying architecture is accessible via public repositories. Ultimately, while pooling legacy hardware offers a viable escape hatch from out-of-memory crashes, users must accept substantial performance trade-offs and recognize that the underlying engine can be replicated without financial investment.
Editorial graphic and headline text by Signal Titans.
Source: This crowdfunded $109 hub wants to pool the RAM in your old PCs, Macs, and phones to run bigger AI models — but the engine underneath is software you can already get for free · GitHub – trademav/ramdeck-core-public: Pool RAM across old PCs, Macs, and Android devices into a unified local AI cluster · GitHub
SIGNAL TITANS
Leave a Reply