Same Models Tested Across Strix Halo, RTX 3090, and RTX 5070
Ran the same models across Strix Halo, RTX 3090, and RTX 5070 because I wanted my own numbers
C_Coffie published 55 local inference benchmark runs across Strix Halo, RTX 3090, RTX 5070, five backends, and 0.35B to 35B-A3B models; RTX 5070 beats RTX 3090 on models fitting 12GiB, while RTX 3090 leads in the 14–31B band that exceeds 12GiB but fits 24GiB.
Why it matters: A user ran 55 local inference tests across three GPUs and five backends, and reports an RTX 5070 decoding faster than an RTX 3090 within 12GB of VRAM.