The RTX 3090 is three generations old, costs under $900 used, and still fits 34B models that a new $500 GPU cannot touch. For LLM inference, VRAM is the hard constraint — and 24GB at ~$820 is the best value on the market in 2026. But the used GPU market has landmines: mining-worn cards, dead VRAM chips, and sellers who know you can’t easily tell the difference before you buy.
This guide gives you the tools to buy one safely.
NVIDIA GeForce RTX 3090
24GB GDDR6X24GB GDDR6X at ~$820 used. Runs CodeLlama 34B, Qwen 2.5 32B, and every 7B/13B model — the same model capacity a new $2,200 RTX 4090 offers, at well under half the price.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
Why 24GB still matters for LLM in 2026
VRAM is a filter, not a preference. A model either fits in VRAM or it doesn’t — and the boundary between “fits” and “doesn’t fit” falls squarely at the 24GB mark for 30B+ models.
- 7B models (Q4_K_M): need ~4.5GB — runs on almost anything
- 13B models (Q4_K_M): need ~8GB — fits an RTX 4060 Ti 8GB, barely
- 34B models (Q4_K_M): need ~20-22GB — requires 24GB VRAM
- 70B models (q4_K_M): a 43GB download — requires dual 24GB cards or a 48GB workstation GPU
For anyone running CodeLlama 34B, Qwen 2.5 32B, or Yi-34B locally, the RTX 3090 is the cheapest single GPU that actually fits the model. A new RTX 5070 Ti (16GB) cannot do it. A new RTX 5080 (16GB) cannot do it. The 3090’s 24GB is the threshold card at the lowest price.
On typical 34B Q4_K_M inference, community benchmarks show a 3090 producing roughly 12-18 tok/s — slow compared to the RTX 4090’s 20-25 tok/s, but well above the ~8 tok/s threshold most people consider interactive. For a full speed breakdown, see RTX 4090 vs 3090 for LLM.
VRAM capacity memory bandwidth Specs are manufacturer figures. Bar lengths are scaled independently per metric.
Price tiers: what to pay and what to avoid
| Price | Signal | Verdict |
|---|---|---|
| Under $600 | Too low — suspect dead VRAM, damaged card, or scam | 🔴 Red flag |
| $600–$699 | Possible mining-heavy card or cosmetic damage — needs heavy scrutiny | 🟡 Caution |
| $820–$900 | Healthy range for a used 3090 with normal wear | 🟢 Target zone |
| $900–$999 | Still reasonable if from a reputable seller with receipts | 🟡 Borderline |
| $1,000+ | Overpaying — well above the healthy range for a card with no warranty | 🔴 Walk away |
Set a ceiling of $900. If you’re patient, quality 3090s appear regularly in the $750-850 range. Cards under $650 almost always have a reason for the discount.
Mining wear vs gamer wear — how to tell the difference
Both mining and gaming cards can be fine to buy. The concern is not the use case — it’s the intensity and conditions of that use.
Mining wear signals (in photos):
- Thermal pads look freshly replaced (miners often replace pads — this is actually good)
- Heatsink fins have no dust (miners keep their rigs clean to manage temps)
- PCB has a slight yellow tinge near VRMs from sustained heat
- Mounting bracket shows no scratches (mining cards rarely get swapped between systems)
- Backplate shows slight bowing — a sign of sustained thermal expansion/contraction cycles
Gamer wear signals (in photos):
- Heavy dust accumulation in heatsink fins
- Scratches on bracket from repeated install/removal
- Original thermal paste (never replaced) — look for grey dried paste at the edges
Neither type is inherently bad, but a mining card that ran 24/7 for 18+ months at 250W+ has more total operating hours than most gaming cards. A gamer card with original paste at 5 years may actually have worse thermal compound degradation.
Questions to ask sellers before buying:
- “What was the primary use — gaming, mining, or professional work?”
- “How many hours of total runtime, roughly? Any way to check?”
- “Has the thermal paste or pads been replaced? When?”
- “Does the card throttle under load? Any driver crashes?”
- “Will you accept a return within 14 days if I find a hardware defect after testing?”
A seller who answers these questions confidently and offers a return window is a better signal than any photo.
Inspection checklist before you commit
Run these checks within the first 48 hours — before your return window closes.
Step 1 — Install and verify with GPU-Z
- Open GPU-Z immediately after driver install
- Check VRAM reported: must show exactly 24384 MB
- Check GPU clock speed: should boost to ~1695 MHz under load
- Any VRAM showing as less than 24GB indicates chip failure
Compare what the tool reports against NVIDIA’s own RTX 3090 specification: 24 GB of GDDR6X and 350 W board power. Those two numbers are the quickest way to catch a card that has been rebadged, flashed with a foreign BIOS, or is quietly running with a dead memory module — all of which turn up on the second-hand market and none of which are visible in a listing photo.
Step 2 — VRAM stress test
- Run CUDA-Z or use
python -c "import torch; t = torch.zeros(24000, 1024*1024//4).cuda(); print('VRAM OK')"in a Python env with CUDA - Alternatively, load a large model in Ollama:
ollama run llama3:70b— this will try to allocate ~40GB (will fail gracefully but exercises VRAM access patterns) - Better: run
memtest_vulkanorOCCT GPU Memory Testto fully exercise all VRAM cells
Step 3 — Temperature and throttle check
- Run a 30-minute Ollama inference session on a 34B model
- Monitor with
nvidia-smi dmon -s pct— watch for thermal throttling (clock dropping while temp is above 83°C) - Expected idle temp: 30-45°C. Under load: 70-83°C is normal, above 85°C sustained is a concern
Step 4 — Fan and coil noise check
- Under load, listen for coil whine (high-pitched electrical buzz — varies from imperceptible to irritating)
- Fan noise: one fan bearing rattling is common and cheap to fix; all three fans rattling means the card was run hard without maintenance
- A brief fan stop at low load is normal (zero-RPM mode)
Step 5 — Backplate inspection
- Remove the card and inspect the backplate for bowing (slight curve away from PCB at center)
- Mild bowing (1-2mm) is common and harmless
- Severe bowing suggests the card was run without proper support — check PCB traces under the backplate if possible
Return-window strategy: Buy from sellers offering at least 14 days returns. Ship to a work address or a friend’s address if you’re buying a second card and your package history makes you a target for “item not as described” scams. Complete all testing within the first 72 hours.
Used RTX 3090 vs alternatives
| GPU | VRAM | Tok/s (13B Q4) | Tok/s (34B Q4) | Price | Notes |
|---|---|---|---|---|---|
| RTX 3090 (used) | 24GB | ~40 tok/s | ~14 tok/s | ~$820 | Best VRAM-per-dollar, no warranty |
| RTX 4090 (new) | 24GB | ~55 tok/s | ~22 tok/s | ~$2,200 | 57% faster on 34B, warranty |
| RTX 5090 (new) | 32GB | ~90 tok/s | ~40 tok/s | ~$4,900 | Runs 34B comfortably; 70B only at q3_K_S or below |
The 3090 is the only option in this table under $1,000. It fits every model the 4090 fits, at roughly 60% of the speed, for barely more than a third of the price. If you’re debating between a used 3090 and a new 4090, see our RTX 4090 vs 3090 for LLM comparison.
Not ready to buy hardware? Try RunPod for RTX 3090 cloud access→Common mistakes when buying a used 3090
Skipping the VRAM stress test. Dead VRAM cells are the most common failure mode on used 3090s. They often don’t show up in normal use — only under sustained 24GB load. Run the test before the return window closes, not after.
Forgetting the PSU when you buy the second one. A single 3090 is fine on most 750W units. Two draw about 700W between them, which puts a 1000W supply at its limit and makes 1200W the sensible size — see the PSU guide for dual-GPU LLM builds before you commit to the pair.
Buying without a return option. Facebook Marketplace deals with no returns are high-risk. If a seller refuses any return policy, you’re betting $800+ on their honesty. The risk-adjusted price of a return-eligible purchase from eBay or Craigslist with a verifiable seller is almost always worth any premium.
Paying $1,000+ because the listing says “lightly used.” Every listing says “lightly used.” Price discipline matters more than seller claims. If you see a 3090 over $950, compare it against used 4090 listings before committing.
Ignoring coil whine. Coil whine doesn’t affect performance, but if it bothers you, there is no fix short of card replacement. Test under load before finalizing, especially if you work in a quiet environment.
Which 3090 buyer are you?
Buy a used 3090 at $800-900 if:
- You need 24GB VRAM for 34B models and can’t spend $2,200 on a new 4090
- You’re comfortable doing a basic hardware inspection and have a return window
- You’re building a multi-GPU LLM setup and need two 24GB cards affordably — our dual RTX 3090 guide walks through the full build
Buy a new RTX 4090 instead if:
- You need warranty coverage and don’t want to test hardware
- You run 34B models interactively all day and want noticeably faster speed
- Your budget reaches $2,200 and you plan to keep the card for 3+ years
Buy a new RTX 5090 instead if: (the full head-to-head runs the numbers)
- You want 32GB VRAM, which reaches a 70B at q3_K_S on a single card — degraded, but without a second GPU
- Budget isn’t the constraint and you want the best available hardware
Verdict
The used RTX 3090 is the best VRAM-per-dollar GPU available for local LLM in 2026, and that isn’t close to changing. The 24GB threshold matters more than generational speed gains at this price point, and nothing new under $2,200 matches it for model capacity.
Buy one in the $750-900 range, run the inspection checklist within 48 hours, and you’ll have a card that handles 34B models for years. The how much VRAM guide explains exactly why 24GB is the sweet spot if you want to understand the math behind the recommendation. With the 2026 GPU shortage tightening new-card stock, used 3090s have become even more attractive — but also more competitively priced.
NVIDIA GeForce RTX 3090
24GB GDDR6X24GB VRAM at ~$820, runs every model under 34B at Q4. The best entry point into serious local LLM inference.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
Frequently asked questions
Is a used RTX 3090 safe to buy for LLM work?
Yes, if you buy from a seller with a return window and run a VRAM stress test within 48 hours. The most common failure mode is dead VRAM cells, which a stress test catches immediately. A mining-worn card that passes a 30-minute stress test is typically fine for LLM inference.
How do I tell if a used RTX 3090 was used for mining?
Look for freshly replaced thermal pads, clean heatsink fins (miners maintain their rigs), slight backplate bowing, and no installation scratches on the bracket. Ask the seller directly about usage history and whether thermal paste was replaced. Mining use alone doesn’t mean the card is bad — what matters is condition.
What is a fair price for a used RTX 3090 in 2026?
The healthy range is $700-900. Cards under $650 typically have a reason for the low price (damage, dead VRAM, or scam). Cards over $1,000 are overpriced — at that point, a used RTX 4090 at ~$1,100-1,200 becomes competitive.
What models can a 24GB RTX 3090 run?
Any 7B or 13B model at any quantization level, and 34B models at Q4_K_M quantization (20-22GB VRAM needed). It cannot fit 70B models in VRAM — a 70B at q4_K_M is a 43GB download, so it needs a dual-GPU setup.
Is a used RTX 3090 better than a new RTX 5070 Ti for LLM?
For model capacity, yes. The RTX 5070 Ti has 16GB VRAM, which cannot fit 34B models. The used 3090’s 24GB is the deciding factor for anyone running 34B+ models. The 5070 Ti is faster per token on models that fit, but if the model doesn’t fit, speed is irrelevant.