Reddit user runs 35B model on 54GB VRAM from $35 crypto rig
A recent acquisition by a Reddit user underscores the growing viability of repurposing legacy hardware for local artificial intelligence workloads. Community member markpronkin secured a dormant cryptocurrency mining rig from a private garage for 3,000 rubles, approximately thirty-five dollars. The seller was equally uncertain about the equipment's condition and refused a pre-purchase power-on test. Upon inspection, the buyer identified a system equipped with nine Nvidia P106-100 graphics processing units, delivering fifty-four gigabytes of collective video memory at a rate of roughly sixty-five cents per gigabyte. After verifying that all modules powered on correctly, the operator immediately conducted a local inference test. The rig successfully ran the 35-billion-parameter Qwen3.6-35B-A3B model through the Ollama software framework, achieving a generation speed exceeding thirty tokens per second. This deployment marks the first confirmed local execution of a model of this magnitude on the discontinued 2017-era P106-100 architecture. The successful test highlights a shifting dynamic in consumer machine learning infrastructure. By utilizing discarded cryptocurrency mining components, independent developers can access substantial computational resources at negligible cost. The operation validates the continued relevance of older GPU architectures for contemporary large language model inference, reinforcing a broader industry trend toward decentralized, cost-efficient AI experimentation.
