llamaperf

RTX 3090 vs RTX 5090

For running local LLMs · 150 reports across 15 models

Side A
RTX 3090
Vendor
nvidia
VRAM
24GB
Memory
Discrete
Side B
RTX 5090
Vendor
nvidia
VRAM
32GB
Memory
Discrete

Highest reported generation speed by model family

These maxima come from different community setups. Model size, quantization, GPU count, offloading and concurrency may differ. They do not establish which GPU is faster for the same workload. Check the comparison conditions →

ModelRTX 3090RTX 5090
Qwen3.8up to 2400B382.0n=51220.0n=35
DeepSeek V4 Flashup to 285B60.0n=9120.0n=10
Qwen3.6up to 35B170.7n=13215.1n=6
Museup to 30B94.0n=1253.0n=4
Gemma 4up to 31B578.0n=4
GLM-5.2up to 744B7.3n=2n=1
GLM-5.3up to 321B55.0n=224.5n=1
Qwen2.5up to 32B70.0n=2n=1
Qwen3up to 235B7.5n=1159.0n=1
DeepSeek R12.0n=1
DeepSeek V3n=1
DeepSeek V4up to 284B15.0n=1
GLM-4.5-Airup to 106Bn=1
Kimi K2.5n=1
Qwen3.5up to 122Bn=1

More comparisons