Can I add a second GPU to my desktop later to double my VRAM?
Yes—for running large language models (LLMs), adding a second GPU effectively pools your VRAM together. Unlike PC gaming, where multi-GPU setups (SLI and CrossFire) are dead, AI engines like Ollama, LM Studio, and llama.cpp automatically divide a large model's layers across multiple graphics cards.
Adding a second 24GB RTX 3090 or RTX 4090 gives you 48GB of combined VRAM, enough to load 70B models that cannot fit onto a single consumer card.
Before planning a future dual-GPU build, keep four practical rules in mind:
- It doubles your capacity, not your speed. When a model is split across two GPUs (known as layer splitting), each card holds a slice of the model and passes activations to the next over the motherboard. Expect roughly 90% of a single card's generation speed—a 5–10% transfer overhead—rather than 2x faster output. Any smaller model that already fits on one card will actually run slightly slower if split across two.
- You do not need an NVLink bridge. Many builders assume a physical
NVLinkconnector is required to pool VRAM. NVIDIA droppedNVLinkfrom every consumer card after the RTX 3090—neither the RTX 4090 nor the RTX 5090 includes it. Modern local AI tools split layers directly across standard motherboardPCIeslots with no bridge hardware required. - You can mix different cards (with a catch). Tools built on
llama.cpp—including LM Studio and Ollama—let you pair mismatched NVIDIA cards, such as a 24GB RTX 3090 alongside a 12GB RTX 3060 for 36GB of total VRAM. Your overall token speed, however, will be bottlenecked by whichever card is slower. - Check your motherboard, case, and power supply first. Two modern graphics cards are thick, hot, and power-hungry. If you want room to drop in a second card later, make sure your desktop has:
- A 1,000W–1,200W+ power supply capable of handling transient power spikes from both cards at once.
- A full-size
ATXmotherboard with two full-lengthPCIeslots spaced at least 3 to 4 slots apart so the top card's fans are not choked for air. - A roomy, high-airflow case so neither card thermal-throttles during sustained inference runs.
Starting with a single 24GB GPU today and adding a second one later is one of the smartest, most modular upgrade paths in local AI—as long as you buy a big enough case and power supply up front.
Looking for verified hardware rigs?
Check our real-time £/GB VRAM index to find systems optimized for local model inference.