Hardware Acquisition via Secondary Markets
A user on the social media platform Reddit has reported the acquisition of a decommissioned cryptocurrency mining rig for approximately $35 USD (3,000 rubles), intended for use in artificial intelligence (AI) workloads. According to a post on the platform that garnered over 1,000 upvotes, the hardware was sourced from a listing on Avito, a Russian classifieds site. The seller had originally listed the equipment for 5,000 rubles (~$60 USD) after discovering a former mining operation in a garage, reportedly stating uncertainty regarding the functionality of the components.
The hardware configuration consists of nine NVIDIA P106 GPUs, each equipped with 6GB of video RAM (VRAM), providing a total of 54GB of VRAM. The user reported that the system was functional upon arrival, requiring only an SSD to operate via a USB boot sequence.
Technical Limitations and Performance Expectations
While the high VRAM capacity offers a potential advantage for loading large language models (LLMs), technical discussions following the post highlight significant performance bottlenecks. Commenters on Reddit noted that the P106 cards are older-generation hardware and are likely to be distributed across limited PCI lanes. This architecture may result in significantly slower data transfer speeds compared to modern enterprise-grade AI hardware.
Users participating in the discussion cautioned that while the total VRAM might allow for the loading of certain models, tasks such as model loading, prefilling, and generating replies could take an extended period. One commenter suggested that the system’s throughput would be “slower than slow” for intensive AI applications compared to contemporary standards.
Optimization and Implementation Efforts
Despite the hardware’s age, the purchaser expressed interest in optimizing the setup for specific AI tasks. The user has sought technical guidance regarding running benchmarks and determining the maximum context window achievable using specific quantization methods, such as Q3_K_M and Q4_K_M.
The user also indicated an intention to utilize advanced software configurations, including requests for specialized bash scripts for model serving and the potential use of tools like vLLM. While the user claims to have successfully utilized AI assistants to assist in recompiling software and setting up configurations, these claims remain unverified by independent technical audits.
The transaction has also drawn criticism from some community members regarding the ethics of negotiating the price down on hardware that may have significant value to specialized users, though the primary focus of the technical community remains on the feasibility of utilizing legacy mining hardware for modern computational intelligence tasks.