ill give u a real answer since most people here just like saying 3090 as it is newer with more vram.

Firstly, it depends on your use case and environment. if you are putting this card in a server rack or having more than 2 in a workstation, the titan rtx will give u better vram per space efficiency as it is only 2 slots while 2 slot 3090s are very rare and expensive for that reason. you can stack 4 titan rtx in a normal workstation case.

If your use case is for single user inference then the titan rtx will save you money and be fast enough for any model that fits in 24gb. for single user inference, you can always use 8 bit 6 bit 4 bit quantized models to save on vram space and increase token speed so spending extra on a 3090 doesnt make sense.

If your use case is multiple user deployment via vllm etc, then there is an advantage in 3090 as it supports bf16 quantization (for vllm u need native hardware support to quantize model weights and kvcache). but honestly the benefits of fp16 to bf16 might not be worth the space inefficiency of 3090 cospared to titan rtx (or the extreme premium u will pay for a 2 slot 3090) depending on your setup. a 4090 will be worth considering more than the 3090 as atleast ada cards support fp8 quantization on model weight and kvcache which halves the weight and kvcache size.

honestly titan rtx costs that much still for a reason. for single user inference, it is pretty good value at 2 slots, 24gb vram and a clean look. only advantage on a 3090 is faster inference (which kinda irrelevant on models you will be running 4 bit 8 bit quantized on 24gb vram) and native bf16 support on vllm (which aint worth the premium ull be paying).