Ollama splits local LLMs across GPUs

Worth doing if your goal is to fit a larger model or run separate models in parallel, not if you expect a clean pooled-VRAM upgrade. The 1070 will add usable memory, but PCIe traffic and the older card’s slower compute will cap throughput.

• –Ollama’s docs say it will load on a single GPU when the model fits there, and only spread across GPUs when it must.
• –Community replies in the thread report that mixed NVIDIA setups work with Ollama and llama.cpp.
• –The 1070’s Pascal-era hardware means it is likely to be the bottleneck on generation speed.
• –Best case: use the 1070 for a second model or auxiliary workload instead of forcing one model to crawl across both cards.
• –For routine local LLM use, a single larger-VRAM GPU is still the cleaner setup.