RAMDeck Pools Hardware Memory to Run Large AI Models
The new hardware hub allows users to combine device memory to offset high DDR5 costs while running complex AI models.
Updated on Oct. 25, 2026 in Computers
Live Poll
Is now a good time to buy hardware to help run large AI models at home?
TradeMAV has launched an Indiegogo crowdfunding campaign for RAMDeck, a hardware hub designed to pool memory across multiple devices to run large AI models. This research-stage hardware uses the llama.cpp RPC backend to split model layers among linked machines.
Why it matters
As DDR5 memory costs remain high, RAMDeck offers a distributed alternative for running large-scale AI models without requiring expensive hardware upgrades. This approach attempts to mitigate the impact of the ongoing global memory supply shortage on local AI development.
The RAMDeck hub achieves inference speeds of 1.92 tokens per second for a 27B model using four devices and 1.37 tokens per second for a 120B model using six devices. These performance metrics rely on the llama.cpp RPC backend, which facilitates model splitting across a local network.
The players
TradeMAV
A developer of orchestration software and hardware hubs that provides automatic discovery and OpenAI-compatible APIs for distributed AI workloads.
Samsung
A global semiconductor manufacturer whose supply projections currently track DDR5 memory availability through 2028.
SK Group
A multinational conglomerate and semiconductor manufacturer projecting potential DDR5 supply constraints lasting until 2030.
The details
RAMDeck functions as an orchestration layer that links computers into a unified memory pool to accommodate large models that would otherwise exceed single-device capacity. It uses a node agent running the llama.cpp RPC server—a remote procedure call protocol used for distributed computing—on each contributing machine to divide and execute model layers. The hub provides automatic discovery features and an OpenAI-compatible API to manage the traffic between the linked nodes.
Timeline
July 2025: DDR5 memory price baseline.
Q4 2025: DDR5 memory prices quadrupled relative to the July 2025 baseline.
November 8, 2026: The crowdfunding campaign closes.
2028: Projected end of DDR5 shortages by Samsung.
2030: Projected end of DDR5 shortages by SK Group.
The Tech Race
RAMDeck positions itself against the rising cost of local AI compute by shifting the burden from high-density RAM to distributed network-linked hardware. This strategy follows a broader trend of leveraging open-source llama.cpp RPC architectures to bypass memory-capacity limitations in current consumer hardware.
Early backers can secure the hub for $109 during the Indiegogo campaign, which requires a $10,000 threshold to reach production. Users should note that current inference speeds are limited by network latency and model size, with performance dropping as the model parameter count increases.
The takeaway
RAMDeck demonstrates an effort to turn idle home hardware into a unified memory cluster to bypass expensive DDR5 upgrades. Watch the progress of the campaign through November 8, 2026, to determine if the hardware hits its production goal.
What happens next
The crowdfunding campaign for RAMDeck concludes on November 8, 2026.
Further reading
For more on the hardware powering local machine learning, visit our Computers section.
Source note: This article includes information reported by TechRadar.
Live Poll
Is now a good time to buy hardware to help run large AI models at home?
