Hello everyone

I am an intern in a software development team.

The company liked the ocr system I developed myself (it uses deep learning models). The projects I made using Gemini api's are the same way.

Unfortunately, the best server I can use in terms of gpu in the office now uses 3050ti. This vram is quite insufficient for both model training, deep learning models and running a local language model. (We want to work locally for data privacy and cost reasons)

As far as I researched 4 x 3090 setup is one of the best budget choices for me. However, I am having trouble finding a motherboard and processor.

For a system with 4 cards (if there is a motherboard with more slots, it would be much better), should all pcie slots be at 16x speed? In some places, there were those who said that the extraction speed would be affected by a maximum of 10% (when x16 and x8 are used together), and in some sources, there were those who said that the speed would drop close to 50%.

According to my usage scenario, do all slots have to be x16 ? Or x16 x8 hybrid use will not give me much performance loss ?

Because motherboards with 4 x16pcie slots are usually server cards. The company does not buy second hand hardware (some new processors are not stocked in my country). These cards are compatible with processors such as the latest threadripper, which creates a high cost that they cannot accept.

There is also an nvlink scenario, if 2 3090s working in 16x and 8x slots are connected with nvlink, will performance loss be prevented? (If this scenario works, I can use i9 processor etc compatible motherboards that host 2 x16 x8 x8).