GeForce RTX 3090 Price & Used Prices: Which Local LLMs Can It Run?
Price alert
We'll email you when the eBay (US) price reaches your target.
NVIDIA's RTX 3090 (2020) is a 24GB flagship gaming card, now mostly used, running Qwen3.8 27B at 4-bit at about 40 tokens/s. Many were mined on.
eBay
GeForce RTX 3090 price history
eBay
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
Full GeForce RTX 3090 specifications
- Memory24 GB · GDDR6X
- Bandwidth936 GB/s
- FP16 / BF16142.3 TFLOPS
- INT8284.7 TOPS
- Bus width
- 384-bit
- Architecture
- Ampere (GA102)
- Cores
- 10,496
- Tensor cores
- 328
- Ecosystem
- CUDA
- Interface
- PCIe 4.0 x16
- NVLink
- 112.5 GB/s
- Power
- 350 W
- Power connector
- 12-pin (Founders Edition)
- Slots
- 3
- Length
- 313 mm
- Type
- Consumer
- Launch date
- Sep 24, 2020
- MSRP
- $1,499
Best GeForce RTX 3090 deals on eBay
Hand-picked deals from trusted eBay sellers. Shop with confidence.
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
UsedEVGA GeForce RTX 3090 FTW3 Ultra Gaming 24GB GDDR6X Graphics Card- READ DESC.Top Rated$1,599Buy on eBay
UsedDell NVIDIA GeForce RTX 3090 24GB GDDR6X Dual Fan OEM 0N78PCTop Rated$1,599Buy on eBay
UsedZOTAC GeForce RTX 3090 24GB GDDR6X Trinity Triple FanTop Rated$1,599Buy on eBay
UsedGIGABYTE Aorus Xtreme Waterforce GeForce RTX 3090 24GB- GPU - Missing Fan ScrewsTop Rated$1,680Buy on eBay
UsedHP NVIDIA GeForce RTX 3090 24GB GDDR6X Triple Fan OEM Non-LHR M24410-002$1,700Buy on eBay
UsedEVGA GeForce RTX 3090 FTW3 Ultra Gaming 24GB GDDR6X GPU 24G-P5-3987-KR$1,700Buy on eBay
What LLMs can a single GeForce RTX 3090 run?
A single RTX 3090 reaches 35B only at 4-bit: 27B to 31B dense at 35 to 40 tokens/s, 30B-class MoE above 130. 8-bit fits only 12B-class.
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
 Gemma 4 E4B8B | 3.8 GB | Runs | 147 | 128K |
 Qwen3.5 9B9.7B | 4.1 GB | Runs | 137.1 | 256K |
| Gemma 4 12B12B | 4.7 GB | Runs | 120.9 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 10.5 GB | Runs | 139.6 | 256K |
| Qwen3.6 27B27.8B | 11.8 GB | Runs | 54.9 | 192K |
| Qwen3.8 27B27.8B | 9.8 GB | Runs | 64.8 | 221K |
 GLM 4.7 Flash30B (3B active) | 11.9 GB | Runs | 150.1 | 198K |
| Gemma 4 31B30.7B | 11.8 GB | Runs | 53 | 143K |
| Qwen3.6 35B A3B36B (3B active) | 12.3 GB | Runs | 160.1 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 E4B8B | 5 GB | Runs | 118 | 128K |
| Qwen3.5 9B9.7B | 5.7 GB | Runs | 105 | 256K |
| Gemma 4 12B12B | 7.1 GB | Runs | 86.1 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 16.9 GB | Runs | 117.5 | 256K |
| Qwen3.6 27B27.8B | 16.8 GB | Runs | 39.7 | 117K |
| Qwen3.8 27B27.8B | 16.5 GB | Runs | 40.4 | 122K |
| GLM 4.7 Flash30B (3B active) | 18.3 GB | Runs | 132.4 | 115K |
| Gemma 4 31B30.7B | 18.3 GB | Runs | 35.8 | 66K |
| Qwen3.6 35B A3B36B (3B active) | 22.1 GB | Runs | 135.4 | 123K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 E4B8B | 8.2 GB | Runs | 77.3 | 128K |
| Qwen3.5 9B9.7B | 9.5 GB | Runs | 67.5 | 256K |
| Gemma 4 12B12B | 12.7 GB | Runs | 51.5 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 26.9 GB | Needs RAM offload | 26.3 | 256K |
| Qwen3.6 27B27.8B | 28.6 GB | Too big | — | — |
| Qwen3.8 27B27.8B | 29 GB | Too big | — | — |
| GLM 4.7 Flash30B (3B active) | 31.8 GB | Needs RAM offload | 22.8 | 198K |
| Gemma 4 31B30.7B | 32.6 GB | Too big | — | — |
| Qwen3.6 35B A3B36B (3B active) | 36.9 GB | Needs RAM offload | 34.7 | 256K |
RunsFits fully in VRAM; speed is estimated from VRAM bandwidth.
Needs RAM offloadMoE only: expert weights sit in system RAM, so speed is estimated from 70 GB/s RAM bandwidth.
Too bigWon’t fit on one card at this quantization, even with RAM offload.
Full calculation rules are on the Methodology page.
What LLMs can multiple GeForce RTX 3090 cards run?
Two 3090s (48GB) suit 48B-class MoE; four (96GB) run Llama 3.3 70B and gpt-oss 120B at 4-bit. NVLink bridges give vLLM a clear two-card boost.
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| Qwen3.8 Flash Next176B (6B active) | 78.9 GB | 4 cards | $7,196 | 96 GB | 111.1 | 1,400 W |
 DeepSeek V4 Flash284B (13B active) | 96.8 GB | 8 cards | $14,392 | 192 GB | 90.9 | 2,800 W |
 Hy3295B (21B active) | 107.2 GB | 8 cards | $14,392 | 192 GB | 44.1 | 2,800 W |
| GLM 5.3 Flash320B (18B active) | 108.7 GB | 8 cards | $14,392 | 192 GB | 74.2 | 2,800 W |
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| Qwen3.8 Flash Next176B (6B active) | 111.3 GB | 8 cards | $14,392 | 192 GB | 94.9 | 2,800 W |
| DeepSeek V4 Flash284B (13B active) | 155.1 GB | 8 cards | $14,392 | 192 GB | 67.9 | 2,800 W |
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| DeepSeek V4 Flash284B (13B active) | 161.9 GB | 8 cards | $14,392 | 192 GB | 66 | 2,800 W |
Full calculation rules are on the Methodology page.
Who it suits
Good fit
- Running 27B to 31B dense models such as Qwen3.8 27B and Gemma 4 31B on one card
They fit at 4-bit, at about 36 to 40 tokens/s
- Running MoE models like Qwen3.6 35B A3B and GLM 4.7 Flash
They use only 3B parameters per token, so one card reaches over 130 tokens/s
- You want minimal software hassle
The CUDA ecosystem is the most complete, and vLLM and ExLlamaV2 work out of the box
- You plan to add a second card later
The 3090 and 3090 Ti are the only 30-series cards that can link two cards with an NVLink bridge
Not a good fit
- Running 70B-class dense models on one card
24GB can't hold them at 4-bit, so you need at least two
- Long full-load sessions when your power bill and noise matter
The card draws 350W, and the rear memory gets very hot in summer
- You need FP8
This chip generation doesn't support FP8 compute, so FP8 models don't run any faster on it
Before you buy
Check rear memory temperature first
Half of the 3090's memory chips (12 of 24) are soldered to the back of the board and cooled only by the backplate, and on mining rigs these chips ran at 100°C to 110°C for months on end. When inspecting a card, run it at full load for ten minutes and watch Memory Junction Temperature in HWiNFO. Holding steady below 100°C is reassuring; approaching 110°C means the thermal pads have aged, and a new set usually brings it down by 10°C to 20°C.
Leave plenty of power supply headroom
Under full load, the 3090 has transient power spikes, which can trip the protection on an undersized or aging power supply and cut power, rebooting the machine. NVIDIA officially recommends 750W for the whole system; even with just this one card, we suggest going to 850W. The Founders Edition uses a 12-pin connector and ships with an adapter; most partner cards use two or three 8-pin connectors.
Will it fit in your case
The Founders Edition is 313 mm long and three slots thick, and many partner cards exceed 320 mm. If you plan to run two cards, first check that your motherboard's two GPU slots are at least three slots apart, or the top card will be starved of air.
Software
Software is mostly worry-free: llama.cpp, Ollama, vLLM and ExLlamaV2 all work out of the box. The one thing to note is that it doesn't support FP8, so when downloading models, pick GGUF at 4-bit or 5-bit, or a 4-bit format such as AWQ or GPTQ, and skip the FP8 versions.
Compared with similar GPUs
GeForce RTX 3090 vs GeForce RTX 4090, Radeon RX 7900 XTX, and GeForce RTX 5070 Ti: price, VRAM, bandwidth, compute and local LLM speed compared.
| Spec | GeForce RTX 3090 | GeForce RTX 4090 | Radeon RX 7900 XTX | GeForce RTX 5070 Ti |
|---|---|---|---|---|
| Memory | 24 GB | 24 GBSame | 24 GBSame | 16 GB33% less |
| Bandwidth | 936 GB/s | 1,008 GB/s8% higher | 960 GB/sSame | 896 GB/s4% lower |
| BF16 / FP16 compute | 142 TFLOPS | 330 TFLOPS132% higher | 123 TFLOPS14% lower | 176 TFLOPS23% higher |
| INT8 compute | 285 TOPS | 661 TOPS132% higher | 123 TOPS57% lower | 352 TOPS23% higher |
| Qwen3.8 27B speed (2-bit) | 64.8 t/s | 69.3 t/s7% faster | 66.3 t/sSame | 62.3 t/s4% slower |
| Largest dense model (4-bit) | 37B | 37BSame | 37BSame | 23B38% less |
| Power | 350 W | 450 W29% higher | 355 WSame | 300 W14% lower |
| GPU interconnect | NVLink 112.5 GB/s | Not supported | — | Not supported |
| Interface | PCIe 4.0 x16 | PCIe 4.0 x16 | PCIe 4.0 x16 | PCIe 5.0 x16 |
| Released | 2020-09 | 2022-10 | 2022-12 | 2025-02 |
| eBay median |
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
GeForce RTX 3090 rental price per hour
On-demand rental of the GeForce RTX 3090 starts at $0.22 per hour (RunPod), about $54/month at 8 hours a day. Buying one (eBay median $1,799) pays off in about 7.8 years, 3.9 years or 1.3 years at 4, 8 or 24 hours a day.
| Option | Price | 4 h/day | 8 h/day | 24 h/day | View |
|---|---|---|---|---|---|
 RunPodRentUpdated 47 min. ago | $0.22/hr | $27/mo | $54/mo | $161/mo | Rent |
 Vast.aiRentUpdated 47 min. ago | $0.35/hr | $43/mo | $85/mo | $255/mo | Rent |
 eBayUpdated 6 hr. ago | $1,799 | Payback ~7.8 yrsPower $8/mo | Payback ~3.9 yrsPower $15/mo | Payback ~1.3 yrsPower $46/mo | Buy |
Break-even uses the US residential average ($0.18/kWh) and 350 W rated power, excluding resale value and the rest of the build.
Rental links are referral links: we may receive a commission or credit if you sign up through them. As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
Amazon.jp
Xianyu