Used RTX 3090 Buying Guide for Local LLM in 2026

Cross-posted from

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): need ~40GB+ — 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

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
$700–$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 — at this price, a used RTX 4090 is ~$1,100-1,200	🔴 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:

1. "What was the primary use — gaming, mining, or professional work?"


2. "How many hours of total runtime, roughly? Any way to check?"


3. "Has the thermal paste or pads been replaced? When?"


4. "Does the card throttle under load? Any driver crashes?"


5. "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

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_vulkan or OCCT GPU Memory Test to 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	~$850	Best VRAM-per-dollar, no warranty
RTX 4090 (new)	24GB	~55 tok/s	~22 tok/s	~$1,600	57% faster on 34B, warranty
RTX 5090 (new)	32GB	~90 tok/s	~40 tok/s	~$2,000	Runs 34B and some 70B at Q4, best new card

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 roughly 55% of the price. If you're debating between a used 3090 and a new 4090, see our and need two 24GB cards affordably — our for years. The tightening new-card stock, used 3090s have become even more attractive — but also more competitively priced.

Read the full guide on Best GPU for LLM — includes our VRAM calculator, GPU comparison table, and live pricing.

Vollständiger Original-Bericht

Ausführliche Details, Code-Beispiele & Hersteller-Stellungnahme auf dev.to.

↗ Original-Artikel auf dev.to lesen