Best Used GPU for Local LLMs: Why a Used RTX 3090 Beats a New RTX 4070
VRAM capacity matters more than raw compute for local LLMs. Used flagship GPUs deliver more VRAM per dollar than any new midrange card. Here is the math and the buying guide.
VRAM per dollar: used vs new
| GPU | VRAM | Used Price | Bandwidth | VRAM/$ | Verdict |
|---|---|---|---|---|---|
| RTX 4090 | 24 GB GDDR6X | ~$1,200 | 1,008 GB/s | 20 MB/$ | Best used overall |
| RTX 3090 | 24 GB GDDR6X | ~$450 | 936 GB/s | 53 MB/$ | Best value |
| RX 7900 XTX | 24 GB GDDR6 | ~$600 | 960 GB/s | 40 MB/$ | Best used AMD |
| RTX 4060 Ti 16 GB | 16 GB GDDR6 | ~$350 | 288 GB/s | 46 MB/$ | Avoid (slow) |
The used RTX 3090 at ~$450 delivers 53 MB of VRAM per dollar - more than 2.5x the RTX 4090 and nearly 3x a new RTX 4060 Ti. For LLMs where VRAM is the bottleneck, this is the metric that matters most. The 3090 also offers 936 GB/s bandwidth, which is 3x the RTX 4060 Ti and enough for comfortable 35B model inference on llama.cpp and Ollama.
What 24 GB used buys that 16 GB new cannot
A new RTX 4070 Ti Super at $800 gives you 16 GB with a warranty. A used RTX 3090 at $450 gives you 24 GB with CUDA. The 50% more VRAM opens up Mixtral 8x7B, Qwen 32B, and Command R 35B at Q4 - models that simply do not fit in 16 GB. For local LLM use, the used card is the better tool.
Used GPU buying checklist
1. Test VRAM
Run CUDA memtest for 15+ minutes. Bad VRAM shows errors quickly. This is the most important test for LLM use.
2. Check thermals
Run a 30-minute stress test. RTX 3090/4090 should stay under 95C hotspot. Higher = bad thermal paste or fan issues.
3. Inspect physically
Look for bent pins, damaged ports, PCB damage. Verify fan spin is smooth with no rattling.
4. Buy with protection
Use eBay, Swappa, or platforms with buyer protection. Avoid cash-only deals unless you can test in person.
5. Benchmark immediately
Run a known LLM benchmark (e.g., llama.cpp prompt eval + generation). Compare speed to expected values for that GPU.
When used does not make sense
- -You need 32 GB. No used consumer card has 32 GB. The RTX 5090 is the only option.
- -You cannot risk downtime. Used cards have no warranty. If it fails, you buy another one.
- -Your PSU is under 650 W. Used flagships (3090, 4090) draw 350-450 W. Factor in a PSU upgrade cost.
- -You are on Windows with AMD. ROCm Windows support is immature. Use Linux or buy NVIDIA.
For the full GPU lineup at every price point, see Best GPU for Local LLMs. Before buying a used card, use our VRAM Calculator to confirm your target model fits in the VRAM of the card you are considering.
Frequently Asked Questions
Is buying a used GPU safe for local LLM workloads?
Should I worry about VRAM degradation on used cards?
How much should I pay for a used RTX 4090?
End of Document
Reader Discussion
Be the first to add a note to this article.
Please log in to join the discussion.
No comments yet.
Related Articles
Best AMD vs Best NVIDIA GPU for Local LLMs: Where AMD Wins, and Where CUDA Still Controls the Market
Can Your GPU Run It? VRAM Compatibility Checker for 80+ LLMs
Used RTX 3090 vs New Midrange GPU for Local LLMs: Why the 3090 Wins on Value
