Dual RTX 3090 local LLM build
48 GB of VRAM from two used cards: the classic 70B setup.
48 GB: dual used RTX 3090 — $5,443.94 for 48 GB of VRAM (2× NVIDIA GeForce RTX 3090 at $1,889 each, 936.2 GB/s). Fits 24 of our 27 sized models, largest Llama-3.3-70B-Instruct (42.52 GB) at ~17 tok/s estimate — about $320 per tok/s for the whole machine.
Part prices are listings dated 2026-10-15; totals exclude tax and shipping.
Last updated Oct 16, 2026. Speeds are bandwidth-based estimates, not benchmarks. Prices are dated snapshots and move daily; "check price" means we have no verified price.
48 GB: dual used RTX 3090 — $5,443.94
Fits 70B at Q4. Two triple-slot cards: verify case and board clearance.
| Part | Item | Price | Source |
|---|---|---|---|
| GPU ×2 | NVIDIA GeForce RTX 3090 (used) | $1,889.00 each | listing · via |
| CPU | AMD Ryzen 5 9600X | $172.00 | listing · via |
| Motherboard | MSI X870E GAMING PLUS WIFI AM5 AMD X870E SATA 6GB/S 5G LAN DDR5 M.2 Wi-Fi 7 include USB 40 | $189.99 | listing · via |
| RAM | Team T-Force Delta RGB 64GB (2 x 32GB) 288-Pin PC RAM DDR5 5200 (PC5 41600) Desktop Memory | $919.99 | listing · via |
| Storage | ADATA LEGEND 710 M.2 2280 2TB PCI-Express 3.0 x4 3D NAND Internal Solid State Drive (SSD) | $219.99 | listing · via |
| PSU | Rosewill CMG5 1200W, ATX 3.0 & 3.1 Compatible, 80+ Gold 1200W Full Modular Gaming Power Su | $99.99 | listing · via |
| Case | NZXT S5 ATX mid-tower | $49.99 | listing · via |
| Cooler | Thermaltake UX400 ARGB (AM5 listed) | $13.99 | listing · via |
| Total (prices dated 2026-10-15) | $5,443.94 | ||
48 GB VRAM. Largest models that fit, with this build's bandwidth of 936.2 GB/s:
| Model | File | Needs | Speed estimate | $ per tok/s |
|---|---|---|---|---|
| Llama-3.3-70B-Instruct | 42.52 GB | 44 GB | ~17 tok/s | $320 |
| DeepSeek-R1-Distill-Llama-70B | 42.52 GB | 44 GB | ~17 tok/s | $320 |
| Qwen2.5-Coder-32B-Instruct | 19.85 GB | 21.4 GB | ~36 tok/s | $151 |
| DeepSeek-R1-Distill-Qwen-32B | 19.85 GB | 21.4 GB | ~36 tok/s | $151 |
| Qwen3-32B | 19.76 GB | 21.3 GB | ~36 tok/s | $151 |
$ per tok/s is the whole machine's price divided by the estimated speed. Full part list and sources: 48 GB: dual used RTX 3090.
How these compare with our other builds
| Build | VRAM | Total | Models that fit | Largest that fits | Qwen3-32B estimate |
|---|---|---|---|---|---|
| 16 GB starter: Radeon RX 9070 | 16 GB | $1,744.44 | 17 of 27 | Mistral-Small-3.2-24B-Instruct-2506 | doesn't fit |
| 16 GB NVIDIA: RTX 5060 Ti 16 GB | 16 GB | $1,834.44 | 17 of 27 | Mistral-Small-3.2-24B-Instruct-2506 | doesn't fit |
| 24 GB: one used RTX 3090 | 24 GB | $2,933.45 | 22 of 27 | Qwen2.5-Coder-32B-Instruct | ~36 tok/s |
| 32 GB: two Radeon RX 9070 | 32 GB | $3,065.92 | 22 of 27 | Qwen2.5-Coder-32B-Instruct | ~25 tok/s |
| 48 GB: dual used RTX 3090 | 48 GB | $5,443.94 | 24 of 27 | Llama-3.3-70B-Instruct | ~36 tok/s |
Speed calculator estimate
Estimate = 0.76 × bandwidth ÷ GB read per token (Q4 file size, scaled by active/total parameters for MoE). Fit = file size + 1.5 GB headroom ≤ total VRAM. Multi-GPU: VRAM adds, speed uses one card's bandwidth (layer split). Short context only. How this works
Questions people ask
What does this build cost?
The 48 GB: dual used RTX 3090 build comes to $5,443.94 for 48 GB of VRAM. Every part is a specific listing dated 2026-10-15; totals exclude tax and shipping, and prices move daily.
What models fit in 48 GB?
24 of our 27 sized models fit the 48 GB: dual used RTX 3090 build, the largest being Llama-3.3-70B-Instruct at 42.52 GB, an estimated 17 tok/s. Against the whole machine's price that is about $320 per tok/s.
Does it run a 70B model?
Yes. Llama-3.3-70B-Instruct is 42.52 GB at Q4_K_M and needs about 44 GB, which fits 48 GB. Estimated speed: about 17 tok/s.
Would one bigger card be cheaper than 2?
We can't say from our price data: the single cards with at least 48 GB (NVIDIA RTX A6000 and NVIDIA RTX 6000 Ada Generation) have no dated price. What we do know is that the 2 NVIDIA GeForce RTX 3090 cards here are $1,889 each, $3,778 of the $5,443.94 total.
Related
- What the NVIDIA GeForce RTX 3090 can run
- What GPU you need for Llama-3.3-70B-Instruct
- What GPU you need for Qwen3-32B
- Can the GeForce RTX 3090 run Qwen2.5-Coder-32B-Instruct?
- Can the GeForce RTX 3090 run Qwen3-32B?
- Build: 16 GB starter: Radeon RX 9070 — $1,744.44, 16 GB
- Build: 16 GB NVIDIA: RTX 5060 Ti 16 GB — $1,834.44, 16 GB
- Build: 24 GB: one used RTX 3090 — $2,933.45, 24 GB
- Build: 32 GB: two Radeon RX 9070 — $3,065.92, 32 GB
- Guide: How much VRAM do you need for a 70B model?
- Guide: RTX 3090 vs RTX 5090 for local LLMs
- Guide: Strix Halo vs a GPU for local AI