Head to head · October 2026
RTX 5080 vs RTX 5090: which is better for local AI?
16GB against 32GB, 960 against 1,792 GB/s. What each runs, how fast, and which one to buy.
The verdict
The RTX 5090 has 16GB more memory (32GB vs 16GB), so it runs larger models, such as Qwen3.6 35B-A3B, without offloading. The RTX 5090 generates text about 67% faster thanks to 1,792 GB/s of memory bandwidth versus 960 GB/s. The RTX 5080 costs about $2,700 less ($1,600 vs $4,300).
- Get the RTX 5080 if you want the most for your money or you care about power draw, heat, and noise.
- Get the RTX 5090 if you want to run bigger models or use longer context or speed matters most.
We may earn a commission when you buy through links on this page. It never changes what we recommend. How we make money
Performance
Model by model
Estimated single-user generation speed for popular models, at the best quantization each card can hold with 8K context.
| Model | RTX 5080 | RTX 5090 |
|---|---|---|
| Llama 3.1 8B Instruct | ~83 tok/sQ8_0 | ~145 tok/sQ8_0 |
| Qwen3 14B | ~68 tok/sQ5_K_M | ~84 tok/sQ8_0 |
| gpt-oss-20b | ~130 tok/sMXFP4 Fits, tight | ~190 tok/sMXFP4 |
| Qwen3 30B-A3B (2507) | ~56 tok/sQ4_K_M Partial offload | ~175 tok/sQ6_K |
| Gemma 3 27B | ~55 tok/sQ3_K_M Fits, tight | ~60 tok/sQ6_K |
| Qwen3 32B | ~5.4 tok/sQ4_K_M Partial offload | ~58 tok/sQ5_K_M |
| Llama 3.3 70B Instruct | ~1.4 tok/sQ4_K_M Partial offload | ~2.7 tok/sQ4_K_M Partial offload |
| gpt-oss-120b | Won't fit | ~25 tok/sMXFP4 Partial offload |
Estimates from our bandwidth model; see methodology. Measured results, where we have them, are on each GPU's page.
Capacity
Models only one of them can run
Only on the RTX 5080
Nothing extra: everything that runs on this card also runs on the other one.
Best models for the RTX 5080Specifications
Spec by spec
| Spec | RTX 5080 | RTX 5090 |
|---|---|---|
| Memory | 16GB GDDR7 | 32GB GDDR7● |
| Bandwidth | 960 GB/s | 1,792 GB/s● |
| Power | 360W● | 575W |
| Price | ~$1,600● | ~$4,300 |
| Price per GB | $100● | $134 |
| Architecture | Blackwell | Blackwell |
| Released | 2025 | 2025 |
| Software | CUDA, Vulkan | CUDA, Vulkan |
Our take
Bottom line on each
Fast, but 16GB is the limit, and the RTX 5070 Ti runs the same models nearly as quickly for less.
The fastest consumer card for local LLMs, but shortage pricing has more than doubled its cost.
Frequently asked questions
Is the RTX 5080 or RTX 5090 better for running LLMs?
The RTX 5090 has 16GB more memory (32GB vs 16GB), so it runs larger models, such as Qwen3.6 35B-A3B, without offloading. The RTX 5090 generates text about 67% faster thanks to 1,792 GB/s of memory bandwidth versus 960 GB/s. The RTX 5080 costs about $2,700 less ($1,600 vs $4,300).
How much faster is the RTX 5090 than the RTX 5080 for AI?
About 67% faster at generating text in our estimates, because token generation is limited by memory bandwidth (1,792 GB/s vs 960 GB/s). Prompt processing depends more on compute and can differ by more.
Can the RTX 5080 and RTX 5090 run the same models?
No. The RTX 5090 fits models the RTX 5080 can only run with CPU offload, including Qwen3.6 35B-A3B and Gemma 4 31B.
Keep comparing