In the end, the RTX 3090 is still better at ML workloads than every card other than the RTX A6000 or the A100. Although the 3090 theoretically falls behind the V100, Titan V and Titan RTX in terms of pure mixed-precision matrix math, in practice the 3090 still performs better on almost all workloads.

So maybe not 100% designed for ML, but better than anything else out there unless you want to sell a kidney.