RTX 3090 24 GB vs RTX 5080 16 GB for local AI
The RTX 3090 24 GB has 8 GB more memory, and that decides it for local AI: it runs a better file (or runs at all) for 32 of 49 models, and 43 run at 8-bit or better against 25. The RTX 5080 16 GB does have hardware FP8, so for models that fit on both it can be faster per step.
RTX 3090 24 GB
- VRAM
- 24 GB
- Memory
- GDDR6X
- Bandwidth
- 936 GB/s
- Architecture
- Ampere
- FP8 compute
- No
- Launch price
- $1,499
VS
RTX 5080 16 GB
- VRAM
- 16 GB
- Memory
- GDDR7
- Bandwidth
- 960 GB/s
- Architecture
- Blackwell
- FP8 compute
- Yes
- Launch price
- $999
01Model by model
| Model | RTX 3090 24 GB | RTX 5080 16 GB | ||
|---|---|---|---|---|
| image models | ||||
| FLUX.1 dev | Runs well Q8_0 | Runs well FP8 | ||
| FLUX.1 schnell | Runs well Q8_0 | Runs well FP8 | ||
| FLUX.1 Kontext | Runs well Q8_0 | Runs well FP8 | ||
| FLUX.1 Krea | Runs well Q8_0 | Runs well FP8 | ||
| FLUX.1 Fill | Runs well Q8_0 | Runs well Q8_0 | ||
| FLUX.2 dev | Runs Q4_K_M | Offload only Q2_K | ||
| FLUX.2 klein 9B | Runs well 16-bit | Runs well FP8 | ||
| FLUX.2 klein 4B | Runs well 16-bit | Runs well 16-bit | ||
| Krea 2 | Runs well Q8_0 | Runs well FP8 | ||
| Qwen-Image | Runs well FP8 | Runs Q4_K_M | ||
| Qwen-Image-Edit | Runs well FP8 | Tight Q3_K_M | ||
| Qwen-Image 2.1 | Runs well 16-bit | Runs well Q8_0 | ||
| Z-Image Turbo | Runs well 16-bit | Runs well 16-bit | ||
| Z-Image | Runs well 16-bit | Runs well 16-bit | ||
| Ideogram 4 | Runs well Q8_0 | Runs Q4_1 | ||
| Boogu-Image | Runs well 16-bit | Runs well FP8 | ||
| ERNIE-Image | Runs well 16-bit | Runs well Q8_0 | ||
| HiDream-O1 | Runs well 16-bit | Runs well FP8 | ||
| Mage-Flow | Runs well 16-bit | Runs well 16-bit | ||
| Lumina 2.0 | Runs well 16-bit | Runs well 16-bit | ||
| HiDream-I1 Full | Runs well Q8_0 | Runs Q5_K_M | ||
| HiDream-I1 | Runs well Q8_0 | Runs Q5_K_M | ||
| SD 3.5 Large | Runs well 16-bit | Runs well Q8_0 | ||
| SD 3.5 Medium | Runs well 16-bit | Runs well 16-bit | ||
| Chroma1-HD | Runs well 16-bit | Runs well FP8 | ||
| SDXL | Runs well 16-bit | Runs well 16-bit | ||
| Illustrious / Pony | Runs well 16-bit | Runs well 16-bit | ||
| SD 1.5 | Runs well 16-bit | Runs well 16-bit | ||
| HunyuanImage 2.1 | Runs well Q8_0 | Runs Q4_K_M | ||
| video models | ||||
| Wan 2.1 14B | Runs well Q8_0 | Runs Q5_K_M | ||
| Wan 2.1 1.3B | Runs well 16-bit | Runs well 16-bit | ||
| Wan 2.1 I2V 480P | Runs well Q8_0 | Runs Q4_K_M | ||
| Wan 2.1 I2V 720P | Runs well FP8 | Tight Q3_K_M | ||
| Wan VACE 14B | Runs well Q8_0 | Tight Q3_K_S | ||
| Wan 2.2 T2V | Runs well Q8_0 | Runs Q5_K_M | ||
| Wan 2.2 I2V | Runs well Q8_0 | Runs Q5_K_M | ||
| Wan 2.2 5B | Runs well 16-bit | Runs well 16-bit | ||
| Wan 2.2 Animate | Runs well FP8 | Tight Q3_K_M | ||
| Wan Animate 2 | Runs well Q8_0 | Tight Q3_K_M | ||
| Wan 2.2 S2V | Runs well FP8 | Tight Q2_K | ||
| SCAIL-2 | Runs well Q8_0 | Tight Q3_K_M | ||
| HunyuanVideo 13B | Runs well Q8_0 | Runs Q6_K | ||
| LTX-Video 13B | Runs well Q8_0 | Runs Q6_K | ||
| HunyuanVideo 1.5 | Runs well 16-bit | Runs well FP8 | ||
| LTX-2 | Runs Q6_K | Tight Q3_K_M | ||
| LTX-2.3 | Runs Q6_K | Tight Q3_K_M | ||
| LTX-2.5 | Runs Q6_K | Tight Q2_K | ||
| MiniMax H3 | Tight Q3_K_M | Offload only Q4_K_M | ||
| MiniMax H3 Pruned | Runs Q6_K | Tight Q3_K_M | ||
Highlighted: the GPU that runs a better file for that model. Launch prices are the maker’s original list prices, not today’s street prices. Calculated from real file sizes; how the numbers work.
02Score
43run well on RTX 3090 24 GB
25run well on RTX 5080 16 GB
32better on RTX 3090 24 GB
0better on RTX 5080 16 GB