Is the RTX 4090 really worth 2.7x the price of a used 3090 for running Ollama? Both have 24GB VRAM, both run the same models. The price gap is $820 vs $2,200 — and it widened during 2026, because the 3090 barely moved while the 4090 climbed. So what are you actually paying for?
Quick answer: The RTX 4090 is 20-40% faster than the RTX 3090 for Ollama inference, but the used 3090 at roughly a third of the price is the better value for most local LLM users.
Check NVIDIA GeForce RTX 3090 on Amazon→Buy on Shopee SG→Who this is for
You want 24GB VRAM for running 13B-34B models on Ollama and you’re deciding between a new RTX 4090 ($2,200) and a used RTX 3090 ($820). Both handle the same model sizes — this is purely a speed-vs-value decision.
Head-to-head benchmarks
| Model | RTX 3090 (tok/s) | RTX 4090 (tok/s) | 4090 advantage |
|---|---|---|---|
| Llama 3 8B (Q4) | ~55 tok/s | ~65 tok/s | +18% |
| CodeLlama 13B (Q4) | ~35 tok/s | ~40 tok/s | +14% |
| Qwen 32B (Q4) | ~18 tok/s | ~25 tok/s | +39% |
| Yi-34B (Q4) | ~16 tok/s | ~23 tok/s | +44% |
The speed gap widens with larger models because the 4090’s higher memory bandwidth (1,008 GB/s vs 936 GB/s) matters more when shuffling bigger model weights.
VRAM capacity memory bandwidth Specs are manufacturer figures. Bar lengths are scaled independently per metric.
Both cards run every model that fits in 24GB. The 3090 is fast enough for comfortable chat on 7B-13B models, and even 34B at Q4 is usable at 16-18 tok/s. For full VRAM planning, see our Ollama guide.
When to buy the RTX 4090
- You run 34B models daily and the speed difference adds up
- You also use the GPU for training or image generation (where Ada Lovelace shines)
- You want new hardware with warranty and no used-market risk
- The $1,380 difference doesn’t meaningfully impact your budget
When to buy the used RTX 3090
- You primarily run 7B-13B models where both cards are fast
- You want to save $1,380 and put it toward RAM or storage
- You’re comfortable buying used (check our used GPU guide)
- You might buy a second one later for multi-GPU 70B
Common mistakes to avoid
- Assuming newer = always better. For Ollama inference, VRAM matters most. Both cards have 24GB.
- Ignoring the power difference. RTX 3090: 350W. RTX 4090: 450W. Over a year of heavy use, that’s $50-100 in electricity.
- Buying a 3090 with mining history. Check the card’s condition. Mining wear reduces lifespan. Ask about usage, test under load.
- Forgetting that two 3090s beat one 4090. For $1,640 — less than a single 4090 — you could get two used 3090s with 48GB combined — that runs 70B models.
Final verdict
| Need | Best pick | Price |
|---|---|---|
| Best value | RTX 3090 (used) | ~$820 |
| Best speed | RTX 4090 | ~$2,200 |
| Best for 70B | 2x RTX 3090 (used) | ~$1,640 |
Both cards run the same models at the same quality. The 3090 is 80% of the speed at 50% of the price. For most Ollama users, that’s the smarter buy.