3090, the 24 GB of VRAM will be more relevant than running smaller models even faster on lower capacity card (small models already run fast by nature of being less to copy from memory and compute on).

3090. Models like flux or wanvideo with fp16 text encoders will push your VRAM usage over 20GB. LLMs with functional context (or workflows with multi-model, not modal, tie in) will all break 16GB.

There's no point to performance without context. At its pricepoint, the 5080 is purely a gaming card.

Do you think you can get a 5080 for a 1000? In the Netherland they're going for 1600 euro's minimum, more like 1800. So I would totally get a new 5080 for a 1000 dollars.

Sigh, I knew they were more than $1000 but did not anticipate that much. You are correct, the price seem to go up every time I check.