The higher end M2s just run more chips in parallel. The bandwidth is there to support the GPU.

The latency of LPDDR is higher than desktop RAM, so it relies on the cache to get around that, and the cache is about the same as most x86 CPUs. So there isn't much benefit in the real world.

That’s okay given the size. The mobo can handle 8000 mt/s, and I can swap sticks whenever I want. Even my gpu has a sweet 24gb ram - more than an average m2. And I can do more than one if I need to train anything.