A Reddit user plans to build a local AI configuration with two GTX 1080 Ti cards and 22GB of combined VRAM, targeting a Qwen model with roughly 27 billion parameters. This shows that second-hand consumer GPUs can still offer an accessible entry point, but “it fits” does not mean “it works well.”
What This Is
The Local LLaMA community is discussing how to run large language models locally on a personal computer, meaning the model runs directly on the user's own device rather than through a cloud service. Each GTX 1080 Ti has 11GB of VRAM, for a combined 22GB. In theory, this is enough to hold some 27B models using 4-bit quantization, which compresses model weights at lower precision to reduce VRAM usage.
Industry View
We believe dual 1080 Ti cards are better suited to low-cost learning and lightweight workloads than to production environments pursuing high throughput. These GPUs are old, and beyond VRAM capacity, memory bandwidth, cooling, power consumption, and software compatibility all affect real-world performance. Multi-GPU inference may also run into model partitioning, framework support, and communication overhead. The key caveat is that 22GB is only the theoretical total: actual available VRAM is usually lower, while model size, context length, and quantization format all consume additional memory.
Impact on Regular People
For enterprise IT: This provides a low-cost way to test whether local models are viable, but enterprise deployments still require assessment of stability, concurrency, maintenance, and total cost of ownership.
For individual professionals: Technically capable users can build document-assistant or code experimentation environments with greater privacy and more controllable costs, but debugging and performance barriers remain.
For the consumer market: Second-hand GPUs still have demand, but “supports local AI” does not mean “suited to local AI.” Capacity, speed, and software support need to be compared together.