Head to head

RTX 4090 vs RTX 3090

Both hold 24 GB, so capacity is a wash. The RTX 4090 has more bandwidth at 1008 GB/s, which decides how fast tokens come out once a model fits.

Model (Q4_K_M, 8K)RTX 4090tok/sRTX 3090tok/s
Llama 3.1 8B Runs comfortably 116 Runs comfortably 112
Llama 3.1 70B Won't fit 1.60 Won't fit 1.67
Llama 3.2 3B Runs comfortably 254 Runs comfortably 246
Llama 3.2 1B Runs comfortably 669 Runs comfortably 650
Llama 4 Scout 109B-A17B Won't fit 5.18 Won't fit 5.39
Qwen3 8B Runs comfortably 112 Runs comfortably 108
Qwen3 14B Runs comfortably 65.5 Runs comfortably 63.5
Qwen3 32B Fits, but tight 30.5 Fits, but tight 29.5
Qwen3 4B Runs comfortably 202 Runs comfortably 196
Qwen3 30B-A3B Runs comfortably 118 Runs comfortably 117
Qwen3 235B-A22B Won't fit 3.23 Won't fit 3.36
Qwen2.5-Coder 32B Fits, but tight 30.5 Fits, but tight 29.5
Qwen2.5 7B Runs comfortably 128 Runs comfortably 125
Qwen2.5 72B Won't fit 1.53 Won't fit 1.59
Gemma 3 4B Runs comfortably 215 Runs comfortably 209
Gemma 3 12B Runs comfortably 78.5 Runs comfortably 76.1

Specifications

RTX 4090RTX 3090
Memory24 GB24 GB
Bandwidth1008 GB/s936 GB/s
FP16 compute165 TFLOPS71 TFLOPS
Launch price$1,599$1,499
ArchitectureAdaAmpere