Hardware comparison
NVIDIA RTX 3060 (12GB) vs NVIDIA RTX 5090 (32GB) for local LLMs
How the two stack up for running open-source LLMs locally — memory, bandwidth, price, and how many of the 69 tracked models each one can run.
Scroll sideways to compare both machines. Keyboard: focus the table, then use the arrow keys. Metric names stay visible.
| Hardware | NVIDIA RTX 3060 (12GB) | NVIDIA RTX 5090 (32GB) |
|---|---|---|
| Memory | 12.0 GB | 32.0 GB (better for this comparison) |
| Bandwidth | 360 GB/s | 1792 GB/s (better for this comparison) |
| Price (approx) | $329 (better for this comparison)GPU only · host PC required, sold separately | $4,399GPU only · host PC required, sold separately |
| LLMs it runs | 30 of 69 | 52 of 69 (better for this comparison) |
| Best model it runs | Mistral Small 3.2 24BQ2_K · estimated 26–39 tok/s | Gemma 4 31BQ5_K_M · estimated 50–75 tok/s |
Both prices cover GPU cards only. A host PC is required and its cost is excluded. Prices are approximate catalog estimates; check current configurations and prices before buying.
The recommended models and quants differ. These estimated speeds are not a like-for-like comparison. Estimates use 4,096 context tokens, llama.cpp and one stream.
The NVIDIA RTX 5090 (32GB) runs 22 more of the tracked models (52 vs 30), thanks to its 32.0 GB of memory. The NVIDIA RTX 5090 (32GB) has more memory bandwidth (1792 GB/s). Higher bandwidth can help generation speed, but model, quant, architecture and runtime also affect the estimates.