A comment from Topaz Labs a while ago suggested that they were advised, by AMD, to not use the matrix units/AI processors in RDNA3 until… something. I can’t remember what exactly.
If this is still true, then TVAI won’t be using the Matrix units on RDNA3, and so any performance differences between RDNA2 and RDNA3 will come from generational changes to the general compute part of the GPU, change to core count, and clock speed.
