Dual-GPU Build: 48 GB VRAM for 70B Models
By LocalLLMGear Editorial · Editorial Team · Updated 2026-06-28
We test hardware hands-on and may use AI tools in research — every guide is human-reviewed. Editorial policy.
We may earn a commission from links in this article, at no extra cost to you. Disclosure.
When a single 24 GB card isn’t enough, the most cost-effective way to reach 48 GB of VRAM is two used RTX 3090s. That unlocks 70B models at genuinely good quality — the kind of jump that a single consumer card can’t make. Here’s how to build it without the classic power and compatibility mistakes.
The 30-second answer: Two used RTX 3090s = 48 GB VRAM, often for the price of one new 4090. Pair them with a 1000W+ PSU, a board with two spaced PCIe x16 slots, and good airflow. No NVLink needed for inference.
Why 48 GB unlocks 70B
A 70B model quantized to 4-bit needs roughly 40–48 GB. One 24 GB card forces heavy quantization (quality drops); 48 GB lets you run it at a higher quality level. Two 3090s are the budget path there.
Parts list
Dual-GPU 48 GB rig — key parts (approx. 2026 prices)
| GPU / Option | Price (approx.) | Best for | |
|---|---|---|---|
| 2× RTX 3090 24 GB (used) ★ Our pick | ~$1,600 | 48 GB combined VRAM | Check price → |
| PSU: 1000–1200W 80+ Gold | ~$200 | Headroom for two cards | Check price → |
| Motherboard: 2× PCIe x16, spaced | ~$250 | Airflow between cards | Check price → |
| CPU + 64–128 GB RAM + NVMe + case | ~$700 | No bottleneck, big airflow case |
Ad · "Check price" links are affiliate links. We may earn a commission at no extra cost to you.
Power and cooling — where builds fail
Two 3090s generate real heat. Use a large, well-ventilated case, space the cards (a spaced motherboard or a riser), and undervolt both GPUs — you lose a few percent of speed for a big drop in power and temperature. Don’t cheap out on the PSU.
Software: splitting a model across two GPUs
Frameworks like llama.cpp, vLLM and Ollama split layers across GPUs automatically over PCIe. Our Tutorials & Setup section covers the configuration.
Not ready to build?
Rent a 48 GB (A6000) or larger instance to test 70B before committing to the build:
Test a 70B model on RunPod AdStarting smaller? See Build a local LLM rig under $2,000.