Custom Budget Rig Pairs 12 CMP 170HX GPUs for 768GB VRAM
A Reddit user assembled a custom hardware rig featuring 12 Nvidia CMP 170HX GPUs to pool 768GB of VRAM for running large local LLMs. The entire setup costs less than a single flagship Nvidia RTX 6000 workstation GPU. High VRAM capacity is the biggest hardware bottleneck for running giant open-weights AI models locally, as enterprise GPUs with massive memory remain prohibitively expensive. This project shows how repurposing specialized mining hardware offers a viable, low-cost path for hobbyists and researchers to achieve enterprise-level memory scale. The system runs software like vLLM and llama.cpp to host models such as GLM-5.3 and Qwen variants, and it uses a fiber optic connection for Remote Procedure Call (RPC) memory extension to a secondary machine. While trade-offs include elevated power consumption, heat, and setup complexity, it effectively bypasses traditional hardware pricing constraints.
## BACKGROUND
Large Language Models require substantial VRAM during inference to store billions of weight parameters and context KV caches. NVIDIA CMP (Cryptocurrency Mining Processor) cards were manufactured specifically for crypto mining without standard video outputs, but their high memory bandwidth makes them attractive for multi-GPU AI inference setups using engines like vLLM.