Single RTX 3090 Real-World Test: Daily Usage Data for Local Agents and Image/Video Generation
The author used a single RTX 3090 24GB, swapping VRAM in and out, to run a local agent and ComfyUI image/video generation simultaneously. Measured: LLM about 72 tok/s, SDXL image about 32 seconds, Wan 2.2 video 4-6 minutes. Data comes from 205 real requests and 35 rendering tasks, not synthetic benchmarks. The author concludes that 24GB can only accommodate either the LLM or one render at a time, and recommends using parallel 1 instead of 4.