RTX 3090 24GB ¥7395 Expensive | Generated by AI
Question: Is a used ASUS/Colorful/Galaxy RTX 3090 24GB at ¥7395 expensive?
Answer:
Yes, ¥7395 RMB for a second-hand RTX 3090 24GB is expensive in the current AI workstation market.
For AI use, the RTX 3090’s value is mainly:
- 24GB GDDR6X VRAM
- CUDA support
- good FP16/BF16 inference
- cheap compared with professional cards
But it is a 2020 Ampere gaming card. (Electronics Hub)
A reasonable 2026 used-market range:
| Card | Reasonable used price (China) | Comment |
|---|---|---|
| RTX 3090 basic models | ¥3500-5000 | good AI value |
| RTX 3090 premium (ROG, HOF, Neptune, water cooled) | ¥4500-6000 | collector / better cooling |
| RTX 3090 Ti | ¥5000-6500 | depends |
| ¥7395 | expensive | only makes sense with warranty + excellent condition |
The market is weird because 24GB VRAM is still valuable for local LLMs. Used RTX 3090 listings globally remain high because people want cheap 24GB cards for AI workloads. (Alibaba)
For comparison:
- Your RTX 4070 has 12GB VRAM, much newer and more efficient, but cannot fit many larger models.
- RTX 3090 gives you 2x VRAM, which is often more important for LLM experiments.
- A used MI50 16GB you mentioned before at ~¥680 is a totally different category: cheap VRAM, but no CUDA ecosystem.
For your use case (GPT training, nanoGPT, Qwen inference, agents):
I would not pay ¥7395.
Better targets:
- RTX 3090 24GB: ¥4000-5000
- Two RTX 3090s: around ¥8000-10000 total → 48GB VRAM, much more interesting for LLM experiments
- RTX 4090 24GB: if price gap is not huge, much faster
The only reason I would consider ¥7395:
- ASUS ROG STRIX / Colorful Neptune / HOF
- original box
- temperature logs
- no mining history
- long warranty
Otherwise it is basically paying “AI shortage premium” for an old gaming GPU. (Financial Times)
For you specifically, since you already have RTX 4070 + are experimenting with MI50, I would wait and hunt a ¥4000-ish RTX 3090. The extra ¥3000+ is better spent on another GPU, RAM, PSU, or storage.