Dual RTX 3090 70B Build (48GB)
As of 2026-10-07. Prices are the 2026-09-28 snapshot (weekly snapshot); board numbers are sourced rows only. Two used RTX 3090s give you 48GB, still the cheapest route past 32GB; the 70B guide covers the software side. The card price moved with September's shortage — this build started at ~$2,500 when a used 3090 was
$1,050 and is **$3,000** at the current snapshot, which is why the title no longer carries a dollar figure. The board has 6 sourced runs on the dual-3090 configuration.
The list
| Part | Pick | Approx. price | Note |
|---|---|---|---|
| GPU ×2 | Used RTX 3090 24GB | ~$1,300 each (2026-09-28 snapshot) | Was ~$1,050 in May 2026 — budget from the snapshot |
| NVLink bridge | RTX 3090 NVLink | ~$120–180 | Optional but meaningful for 70B Q4 latency (see caveats) |
| Motherboard | X570 / B550 with 2× PCIe x16 (spaced) | $150–250 | x16/x16 or x8/x8; x8/x8 is fine for inference |
| CPU | 6–8 core recent-gen | $120–180 | |
| RAM | 96GB DDR4/DDR5 | $180–250 | 70B Q4 + KV cache wants system-RAM headroom |
| PSU | 1200W 80+ Platinum | $180–250 | 2×350W under load; don't undersell this |
| Case | Full ATX, horizontal slot spacing | $80–120 | Both cards need airflow room |
| Total | ~$2,900–3,100 |
What it actually runs (board rows, per-row attribution)
- Qwen3.8-27B W4A16 — 182.4 tok/s (imported from LocalMaxxing) — the 27B dense class, ~2.2× the single-3090 Q4 row (81.9 tok/s).
- Nemotron-3.5-Lightning-30B-A3B NVFP4 — 197.0 tok/s (imported) — 30B MoE class.
- Qwen3.6-27B AutoRound INT4 — 240.8 tok/s (vLLM, 200K context) — the best sourced row on this configuration, and small-model-class speed with 48GB of room.
- The headline row: 70B dense at Q4 — ~40GB of weights fits the combined 48GB; per the 70B guide, expect patient-but-usable speeds (single-digit to low-tok/s range on dense 70B) versus the A6000's ~12 tok/s class.
Why dual 3090 instead of one A6000 48GB
| Dual 3090 (~$2,300–2,600, total ~$3,000) | Used A6000 (~$2,500, total ~$3,000) | |
|---|---|---|
| VRAM | 48GB (2×24) | 48GB |
| Bandwidth | 936GB/s aggregate | 768GB/s |
| 70B Q4 speed | Good (P2P-bound) | Slightly faster per-GB, simpler |
| Setup | NVLink/P2P, 2-slot spacing, 1200W | One card, one slot |
| Resale | Two liquid cards, deep market | One card, thinner market |
| 2026 risk | Used-condition roulette ×2 | Professional card, longer life |
The dual wins on price, resale, and raw aggregate bandwidth. The A6000 wins if you value "one card, no mods, no fan-lottery ×2."
Honest caveats
- You are buying two used cards with zero warranty. Test fan curves before paying; ask for a burn-in video. One bad card means the build sits.
- P2P/NVLink is not optional for the 70B use case if you want the best latency — without it, weights split across cards and inter-card traffic goes through the PCIe root complex.
- Power: two 350W cards is a real 1200W bill at peak. Budget the electricity.
- The used-3090 price floor has been rising since May 2026 (VRAM shortage). If the weekly snapshot shows $1,200+, this build creeps past $3,000 and the A6000 comparison flips.
Sources
- Used 3090 pricing: Howard Resource Group (eBay sold, 2026-05) + weekly snapshot
- Software setup: run 70B on used RTX 3090s
- Benchmark rows: RTX 3090 ×2 on the board