This week a post on the LocalLLaMA community gained traction: we noticed a developer daisy-chained a 2017 GTX 1080 Ti with this year's RTX 5070 Ti over gigabit Ethernet, using llama.cpp's RPC feature to combine them into an AI workstation that runs a quantized Qwen 27B. What makes this worth our attention is — those dusty old GPUs sitting in corporate server rooms may finally not be scrap metal anymore.
What this is
llama.cpp is an open-source tool that lets ordinary computers run large language models. This developer used its RPC (Remote Procedure Call) feature to coordinate two GPUs across machines: the 1080 Ti handles the main computation, while the 5070 Ti handles MTP (Multi-Token Prediction, a speculative decoding acceleration), with data shuttling across gigabit Ethernet. At 12k context (how much text the model can "see" at once), output speed hit 36 tokens per second, and the quantized Qwen 27B (compressing the model to fit smaller VRAM) ran smoothly.
Industry view
The local AI community is energized — old hardware is no longer obsolete, but a ticket to low-cost deployment. Open-source models like Qwen and Mistral have matured in quality, and quantization techniques let smaller-VRAM cards run complex tasks.
But sober voices are there too: 36 tokens/s is far below cloud API response speeds; gigabit Ethernet itself is a bottleneck; the developer himself admits that "prefill (processing input) and generation (output text) are mutually exclusive — no middle ground." More critically, MTP acceleration eats 30% more VRAM, which actually turns off users with small cards. You can save money, but don't expect this to truly replace cloud services.
Impact on regular people
For enterprise IT: it's worth taking stock of dusty old GPUs in server rooms — they might serve another three years. Deploying LLMs locally doesn't require a one-time big investment.
For working professionals: the tech-savvy can build a low-cost local AI assistant for handling sensitive documents without data leaving the premises; but most white-collar workers will find using ChatGPT or Wenxin (ERNIE) directly more hassle-free.
For the consumer market: more "all-in-one AI" machines may emerge, bundling old hardware repurposing with local AI and selling them to SMBs.