There haven't been many GPUs missing a 32-bit part of the memory bus [1]. This means that it has either a misbalanced memory configuration (some memory channels having more memory than others) or a cut-down 160-bit memory bus.
However, if we assume that Navi 22 has 384-bit memory bus then the answer is straight-forward - the cut down version has one double-memory controller less, 320-bit in total. This would also mean that every other rumored bus size, like N21 being 256-bit, should also be multiplied by 2. In my opinion that is a more likely explanation than the "infinity cache" that would magically cut the memory bandwidth needs of a huge GPU in half (the cache might still exist but have a different purpose, e.g. to speed up unpredictable memory access patterns of ray tracing BVH traversal).
Please note though that this is just my assumption that might be completely wrong.
[1] Technically, single GDDR6 channel is 32-bit. Looking at the wiki page, it seems there was at the very least a cut-down GeForce 1050 with 96-bit bus, and 1080Ti/2080Ti use 352-bit bus. For AMD, there is a single cut-down Navi-14. Not many in total.
en.wikipedia.org
en.wikipedia.org
