[Official FAQ] Share your experience or follow-up questions below.

You can, but the biggest benefit is usually fitting a larger model, not doubling speed.

• llama.cpp: splits layers across GPUs. You can adjust the proportion with --tensor-split, and mixed GPU models can work. With layer splitting the cards mostly work one after another for a single request, so speed gains are limited.
• Ollama / LM Studio: typically detect multiple GPUs and spread the model automatically. Behavior depends on the version.
• vLLM: supports tensor parallelism, for example --tensor-parallel-size 2. It generally prefers identical GPUs, and the parallel size usually needs to divide the number of attention heads, so 2, 4 or 8 are common. The benefits are clearer with many concurrent requests.

Things to keep in mind

• Interconnect matters. Limited PCIe lanes or no NVLink can reduce tensor-parallel efficiency.
• Two GPUs mean more power draw, heat and PSU requirements. Plan the build carefully.
• Total VRAM is not perfectly additive: each card has its own overhead, and the split may be uneven.
• If a model already fits on one card, a second card will rarely be worth it just for speed.

Practical advice: for personal use of a model that does not fit on one card, llama.cpp layer splitting is a simple option. For serving many users, look at vLLM tensor parallelism.

Do you run a multi-GPU setup? Share your cards, software and experience.