Is 3090 + 5070 & 5060s a good idea?

I have a 5070 Ti and two 5060 Ti (all 16Gb cards).

I planned to add another 5070 Ti giving me two pairs of 32Gb each but the NVidia prices have just jumped by 25% where I am and I've found a 3090 Founders Edition for a good chunk cheaper than the 5070 would cost.

It's 8Gb more VRAM and even slightly higher memory bandwidth, but I've read that mixing Ampere with Blackwell comes with a performance hit in llama.cpp using tensor parallelism. I believe TP isn't possible at all in vLLM with mismatched cards.

I am getting 60 tok/s decode and 1,500 pp out of Qwen3.8-27B-IQ4-XS-MTP in llama.cpp with TP across the 5070 and one of the 5060s. I don't really want to spend a lot of money for 8Gb more VRAM in my main pair but worse performance.

Has anyone got experience mixing similar cards? Would pipeline-parallelism (layer split) be faster with two high memory bandwidth cards? I know I could set the 3090 as the main card and push the K/V cache onto it.