I spent a lot of time researching and testing different methods for that, the only thing that worked was LACT in Linux. Using that I was able to undervolt 100mV and GPU power usage dropped about 10%. On my B450 ITX board with a Ryzen 2400GE CPU the entire system pulls 30w idle from the wall, and about 300w inferencing with VRAM filled. (330w before LACT).
My fan solution ended up being to buy the 80mm 3d-printed shroud off ebay, the fan that came with it was super loud so I switched to an arctic p8 Max, and control it with the motherboard targeting a t-sensor header with the probe attached to the backplate.