RTX 5070 Ti vs RTX 3090 — MiniMax H3 image-to-video benchmark (non-turbo, 20 steps)
retained text
Two LAN ComfyUI render boxes: RTX 5070 Ti (16 GB) vs RTX 3090 (24 GB). Ran MiniMax H3 image-to-video, non-turbo — 22B int8 UNet + 32B Qwen3VL CLIP (nvfp4), 20 full steps, 5 s clip at 704×1280, identical start frame and prompt, int8 video VAE on both, fresh noise seed every pass. The 5070 Ti wins by \~25% — and it won cold too, which surprised me given it has 8 GB less VRAM. H3 at full steps is pure compute: the sampler node alone took 474 s vs 645 s. The 3090's VRAM headroom never mattered; everything fit even on 16 GB once offloaded. Same config both boxes: ComfyUI standalone portable, --fast fp16\_accumulation, torch cu130, default attention (PyTorch SDPA — cuDNN kernels on Blackwell, FA2/mem-efficient on Ampere), no flash-attn. TL;DR: for H3-class video work the 5070 Ti is \~26% faster end to end. The 3090 only makes sense here if your scene genuinely exceeds 16 GB. Caveats: single card each, consumer standalone builds, not a lab test. First warm pair showed 2 s passes — that was ComfyUI's execution cache reusing an identical graph; re-ran with randomized seeds. Numbers above are the real ones.
document fetched 2026-11-14 00:31 · canonical https://reddit.com/r/StableDiffusion/comments/1xvey5t/rtx_5070_ti_vs_rtx_3090_minimax_h3_imagetovideo