Home › GPUs & Macs › NVIDIA RTX 3090 vs NVIDIA RTX 5090
NVIDIA RTX 3090 vs NVIDIA RTX 5090 for local LLMs
Across 7 tests measured on both, the NVIDIA RTX 5090 is faster in 4 and generates a median 1.75× the tokens/s of the NVIDIA RTX 3090.
On Qwen3.8 27B (Q4_K_M, llama-server) the NVIDIA RTX 3090 generates 36.6 tokens/s and the NVIDIA RTX 5090 76.9 tokens/s (measured).
Same model, same quant, same test
| Model | Quant / test | RTX 3090 tok/s | RTX 5090 tok/s | Faster | Sources and run details |
|---|---|---|---|---|---|
| Qwen3.8 27B | Q4_K_M llama-server | 36.6 CUDA | 76.9 CUDA | RTX 5090 +110% | qwen38-mtp RTX 3090 · qwen38-mtp RTX 5090 |
| Qwen3.8 27B | UD-Q4_K_XL llama-server | 34.0 CUDA | 74.3 CUDA | RTX 5090 +119% | qwen38-mtp RTX 3090 · qwen38-mtp RTX 5090 |
| Llama 2 7B | Q4_0 llama-bench | 172 Vulkan | 300 CUDA | RTX 5090 +75% | llama.cpp #10879 RTX 3090 · llama.cpp #15013 RTX 5090 |
| gpt-oss 20B | MXFP4 llama-bench | 162 CUDA | 283 CUDA | RTX 5090 +75% | llama.cpp #15396 RTX 3090 · llama.cpp #15396 RTX 5090 |
| Llama 3.1 8B Instruct | Q4_K_M LocalScore | 96.3 llamafile | 66.3 llamafile | RTX 3090 +45% | LocalScore RTX 3090 · LocalScore RTX 5090 |
| Llama 3.2 1B Instruct | Q4_K_M LocalScore | 270 llamafile | 120 llamafile | RTX 3090 +125% | LocalScore RTX 3090 · LocalScore RTX 5090 |
| Qwen2.5 14B Instruct | Q4_K_M LocalScore | 56.2 llamafile | 41.8 llamafile | RTX 3090 +35% | LocalScore RTX 3090 · LocalScore RTX 5090 |
Generation speed in tokens/s, best measured backend for each chip (shown under the number). Backends can differ between the two chips; check them before drawing conclusions. Measured results only, nothing estimated.
Specs
| NVIDIA RTX 3090 | NVIDIA RTX 5090 | |
|---|---|---|
| Memory | 24 GB | 32 GB |
| Memory bandwidth | 936 GB/s | 1792 GB/s |
All results for the NVIDIA RTX 3090 · All results for the NVIDIA RTX 5090 · More comparisons