TL;DR: A pair of used RTX 3090s costs $2,912–$3,116 as of October 20, 2026 — the same money as one used RTX 4090 at its $2,999 median asking price. The pair buys 48GB and runs Llama 3.3 70B fully resident at 16–32 tok/s; the 4090 buys one fast, simple 24GB card that can’t load the model at all.

	2× Used RTX 3090	1× Used RTX 4090	1× Used RTX 3090
Best for	70B-class models fully in VRAM	Fastest 24GB card + diffusion/fine-tuning	Everything that fits 24GB, half the money
Price (Oct 2026)	$2,912–$3,116 the pair	~$2,500 market avg; $2,999 median ask	$1,456–$1,558
The catch	Dual-GPU is a project: board, PSU, topology	42.5GB of 70B weights won’t fit in 24GB	Same ceiling as the 4090, 12% slower

Honest take: If you’ll actually run 70B-class dense models, buy the two 3090s — no single card near $3,000 can hold them, and that’s the whole decision. If everything you run fits in 24GB, don’t buy either of the $3,000 options: buy one 3090 and keep $1,300, unless you generate images or fine-tune daily, which is the one workload where the 4090’s compute earns its price.

For most of 2026, the pair-of-3090s question was academic — two used 3090s cost meaningfully more than one used RTX 4090. October’s repricing closed the gap. ResalePrices’ fair asking range for a used RTX 3090 is now $1,456–$1,558 across 320 tracked listings ($1,533 average, October 20, 2026), so a pair lands at $2,912–$3,116. The used 4090 sits right inside that window: ResalePrices tracks a ~$2,500 market average across 17 active listings (flat over 90 days), while DealHunter’s October 20 snapshot of 49 mostly-eBay listings shows a $2,831 average and $2,999 median asking price.

Same ~$3,000. Two completely different machines. One is 48GB of pooled VRAM across two slots; the other is 24GB behind one power connector. Which one is right depends almost entirely on one question — what’s the largest model you actually intend to run? Check your exact model-plus-context numbers in the VRAM calculator as you read.

The spec sheet, and the row that decides it

	2× RTX 3090	1× RTX 4090
VRAM	48GB pooled (24+24)	24GB
Bandwidth per card	936 GB/s	1,008 GB/s
Board power	350W × 2	450W
Power connectors	8-pin PCIe (standard)	12VHPWR (melt history)
NVLink	Yes (last consumer card with it)	Dropped on Ada
Motherboard requirement	Two x8+ electrical slots	Any board
Warranty	None, twice	None

Per-card bandwidth differs by 7.7%, and since LLM decode is memory-bandwidth-bound, that’s roughly the single-model speed gap you’d expect — and roughly what shows up in measured results below. The row that actually decides the purchase is VRAM, because model weights either fit or they don’t.

What 48GB buys: the 70B class, resident

Llama 3.3 70B at Q4_K_M is a 42.5GB file. On any single 24GB card — 3090 or 4090, doesn’t matter — it cannot load, and CPU offload gets you 2–5 tok/s, below reading speed. The XiongjieDai llama.cpp benchmark suite shows exactly this: single 3090 and single 4090 both report OOM on 70B Q4_K_M at every context length tested.

On two 3090s, the same model runs fully in VRAM. Measured results span a wide range, and the spread itself is the useful information:

Setup	70B Q4_K_M decode	Source
2× 3090, layer split (default), clean topology	16.3–19.5 tok/s	XiongjieDai suite (16.29 at 1K ctx); computingforgeeks 2026 build (19.5)
2× 3090, tensor split (--split-mode row)	31.9 tok/s	computingforgeeks (single test, 20.8 GiB peak per card)
2× 3090, one card on a chipset x4 slot	single digits, or corrupted output	community rigs in our dual-3090 wiring guide; llama.cpp issue #20052
2× RTX 4090 (~$5,300+ the pair)	19.1 tok/s layer split	XiongjieDai suite

Three things worth reading out of that table. First, the physics: with layer split the cards take turns, so the ceiling is one card’s bandwidth — 936 GB/s ÷ 42.5GB per token ≈ 22 tok/s — and the 16.3–19.5 measurements sit at 74–89% of it, exactly where healthy builds land. Tensor split keeps both cards reading weights simultaneously, which is how 31.9 clears the single-card ceiling. Second, the warning: tensor/row split is topology-sensitive — Hardware Corner’s 3090 testing measured prompt processing dropping roughly 60% with row split on PCIe-only rigs, and issue #20052 documents a dual-3090 machine producing garbage above 2K context because one card hung off a chipset x4 slot. Treat 31.9 as what a correctly wired x8/x8 build can reach, not a default. Third, the comparison that kills the “just buy two 4090s” idea: the 4090 pair decodes 70B at 19.1 tok/s — 17% faster than the 3090 pair’s layer-split number for roughly $2,300 more, because decode follows bandwidth, not compute.

What the pair does not buy is speed on smaller models. In the same suite, Llama 3 8B Q4_K_M runs at 111.7 tok/s on one 3090 and 108.1 tok/s on two — the second card adds capacity, not velocity. If your models fit on one card, the second card is $1,500 of silence.

Sanity-check a build before trusting any of these numbers on your own rig:

$ nvidia-smi --query-gpu=index,name,memory.used,pcie.link.gen.current,pcie.link.width.current --format=csv
index, name, memory.used [MiB], pcie.link.gen.current, pcie.link.width.current
0, NVIDIA GeForce RTX 3090, 21299 MiB, 4, 8
1, NVIDIA GeForce RTX 3090, 21184 MiB, 4, 8

Both cards near 21GB used means the 70B is split and resident. A width.current of 4 on either card means it’s on a chipset slot — that’s the 3-tok/s trap, and no software flag fixes wiring.

What the single 4090 buys

Measured speed on everything that fits. 127.7 vs 111.7 tok/s on 8B Q4_K_M (+14%), and on the GPT-OSS 20B MoE the gap widens to 225 vs 161 tok/s in the llama.cpp community benchmarks, where Ada’s compute helps beyond raw bandwidth. Qwen3.6-27B Q4_K_M (16.8GB) runs resident on both — 40.1 tok/s on a 3090, proportionally faster on the 4090.

Compute workloads the 3090 pair can’t match. Diffusion and fine-tuning are compute-bound, not bandwidth-bound, and they don’t split across consumer cards the way inference does. For daily FLUX/SDXL work or QLoRA runs, one 4090 is roughly twice the card a 3090 is, and hardware FP8 support (Ampere has none) increasingly matters as FP8 checkpoints become the default fast path. We covered the full single-card case in the used 4090 verdict.

A purchase instead of a project. One card works in any PC with a decent PSU. Two cards need a board with two x8-electrical slots (the ProArt-class boards we specced in the $5,000 build run ~$480), a PSU sized for ~700W of GPU load (1200W class, vs 850–1000W for a 4090), a case and airflow plan for two 350W triple-slot cards, and an afternoon of BIOS/IOMMU/topology verification. None of it is hard for a home-labber, but it’s real time and ~$300–$500 of platform cost the single card skips.

One real liability, though. Every used 4090 carries the 12VHPWR connector. NVIDIA acknowledged ~50 melted-connector cases early on, and a reviewer found a card that ran two years with a quietly melted connector. Inspect the 16-pin socket for discoloration before money changes hands. The 3090’s standard 8-pin connectors have no such history — buying two of them is two rolls of the ordinary used-GPU dice (mining history, fan wear), not two rolls of the melt dice.

Power: the pair costs more every month

At the August 2026 US average residential rate of 18.44¢/kWh, two hours of heavy inference a day costs about $7.74/month on the dual-3090 rig (~0.70kW under load) versus $4.26/month on the 4090 (~0.385kW measured draw) — our arithmetic from those rates. Not decisive at 2h/day; it compounds if the box runs 24/7. The standard dual-3090 move is power-limiting both cards to ~280W, which barely touches decode speed (decode never used the full 350W) and takes the pair’s ceiling under 600W.

The decider: what do you actually run?

Be honest about this, because it’s the whole purchase:

• The 24GB menu — everything in our 24GB tier guide: Qwen3.6-27B, Gemma 4 at tight quants, GPT-OSS 20B at long context, every 7–14B at huge context. If this is your ceiling, the right buy is one used 3090 for $1,456–$1,558 — the 4090 does the same list ~8–14% faster for about $1,300 more, which only makes sense with a diffusion or fine-tuning habit.
• The 48GB menu — the 48GB tier: Llama 3.3 70B Q4 resident, 70B-class fine-tunes, 27–32B models at maximum context, or two models loaded at once (a chat model and a coding model serving a local coding stack). If anything here is why you’re spending $3,000, the dual 3090s are the only way to get it at this price — the next cheapest resident-70B CUDA option is workstation-card money.

MoE sizes complicate the picture honestly: GPT-OSS 120B’s weights (~63GB) don’t fit 48GB either, so the pair doesn’t unlock everything — it unlocks the dense-70B class and the long-context mid-size class. Run your actual target through the VRAM calculator before deciding.

What to actually buy

Prices as of October 2026, all verified in the comparison above:

Your situation	The machine	Price	Where
You’ll run 70B-class models and want them resident	2× used RTX 3090	$2,912–$3,116	Check price
Everything fits 24GB + you do daily image gen or fine-tuning	Used RTX 4090	~$2,500–$3,000	Check price
Everything fits 24GB and the GPU’s job is serving tokens	1× used RTX 3090	$1,456–$1,558	Check price
Want to test your 70B workload before committing $3,000	Rented 3090 from $0.07/hr, 4090 from $0.14/hr	pay per hour	Vast.ai

The rent row is worth taking seriously: an evening of testing your actual models on a rented dual-GPU instance costs a couple of dollars and answers the only question that matters — whether 16–32 tok/s on a resident 70B is worth more to you than one simple fast card.

FAQ

Will two 3090s make my 13B model faster? No. Models that fit on one card run at single-card speed — 8B measured 111.7 tok/s on one 3090 and 108.1 on two. The second card buys capacity, not velocity.

Do I need NVLink for the pair? For llama.cpp inference, no — layer split barely touches the link, and the measured numbers above are all PCIe-only. NVLink helps tensor-parallel serving stacks like vLLM, and the 3090 is the last consumer card that supports it (our NVLink vs PCIe breakdown). Don’t pay a big bridge premium for chat inference.

Why not one RTX 5090 instead? A RTX 5090 runs $3,822–$5,000 when you can find one, and its 32GB still doesn’t hold a 42.5GB 70B — you’d be paying more for a card that also can’t run the model class this comparison is about.

Is a used 3090 safe to buy in 2026? It’s a 5–6 year old card with zero warranty and likely mining history — the risks and the inspection checklist haven’t changed. Buy with returns, test under sustained load. The same applies to a used 4090, plus the 12VHPWR socket inspection.

What platform do the two cards need? Two x8-electrical PCIe slots (CPU lanes, not chipset), a 1200W-class PSU, and four slots of physical clearance. The dual-3090 wiring guide covers the topology traps that silently cost half your speed.

Sources

1. ResalePrices — RTX 3090 used market tracking ($1,456–$1,558 fair range, $1,533 avg, 320 listings, Oct 20 2026)
2. ResalePrices — RTX 4090 used market tracking (~$2,500 market average, 17 listings)
3. DealHunter — used RTX 4090 median and 30-day trend ($2,999 median ask, $2,831 avg, 49 listings, Oct 20 2026)
4. BestValueGPU — RTX 4090 price history (~$2,999 used, Oct 2026)
5. XiongjieDai — GPU benchmarks on LLM inference (8B/70B Q4_K_M on 1×/2× 3090 and 4090)
6. ComputingForGeeks — dual RTX 3090 LLM build and 70B benchmarks (19.5 tok/s layer split, 31.9 tensor split)
7. llama.cpp discussion #15396 — GPT-OSS community benchmarks (3090 161 tok/s, 4090 225 tok/s)
8. Hardware Corner — RTX 3090 LLM benchmarks (row-split prompt-processing penalty on PCIe)
9. llama.cpp issue #20052 — dual-3090 chipset-slot corruption
10. TweakTown — NVIDIA on melted 12VHPWR connectors
11. Tom’s Hardware — hidden melted connector found two years later
12. ChooseEnergy — US electricity rates by state (18.44¢/kWh US average)

Recommended Gear

• Used RTX 3090 24GB — $1,456–$1,558 each; two of them are the cheapest resident-70B machine that exists
• Used RTX 4090 24GB — ~$2,500–$3,000; the one-card answer if your models fit 24GB and you also create images or fine-tune

Prices verified as of October 20–21, 2026. Used-GPU prices move weekly; check the trackers linked in Sources before buying.

Get the numbers before you buy

New GPU and mini-PC benchmarks, VRAM thresholds, and price checks — sent when there's something worth acting on, not on a schedule. No spam, unsubscribe anytime.

Email address Subscribe