← sources

GPU showdown: RTX 5070 Ti vs RTX 3090 in ComfyU

source: reddit (reddit:StableDiffusion)published 2026-11-13 14:11ingested 2026-11-14 02:38

retained text

*Processing img 99rnlpjd1e1i1...*

|Benchmark|RTX 5070 Ti (16 GB)|RTX 3090 (24 GB)|Winner|
|:-|:-|:-|:-|
|Krea image – cold load|15.1 s|12.9 s|3090 (+15%)|
|Krea image – warm avg|7.7 s|10.4 s|5070 Ti (+26%)|
|LTX video – cold load|82.6 s|69.9 s|3090 (+15%)|
|LTX video – warm avg|45.1 s|49.4 s|5070 Ti (+9%)|

Takeaways

* Warm/batch work: 5070 Ti wins both — native FP8 tensor cores crush the emulated path on Krea fp8 (26% gap); the gap shrinks to 9% once the model is already int8 (LTX).
* Cold one-offs: 3090 wins everything — 24 GB means no VRAM offload, faster model loads.
* Practical routing: batches on the 5070 Ti, big-VRAM models on the 3090.
* Caveat: single 3090 vs single 5070 Ti, consumer ComfyUI, not a controlled lab test.

document fetched 2026-11-14 00:31 · canonical https://reddit.com/r/StableDiffusion/comments/1xv8w8l/gpu_showdown_rtx_5070_ti_vs_rtx_3090_in_comfyu