The issue with a RNG and other hardware of similar character is not the bandwidth, it’s the latency: how long between first request and first result. That needs to be no more than 4-5 cycles, and uniformly consistent for every use. Otherwise the hardware can exist, but it needs to be accessed in a different way or the whole machine suffers whether the RNG is used or not. So the right place for a RNG is as a side hunk of hardware, not reached by the usual operator interface. The timing, bot latency and bandwidth, would be as good as if it were an operation, but without the penalty to machine performance of other code.
Similarly, the belt cannot be indexed or have an address taken. Again, even the possibility of doing do would cut everybody’s performance 2X or more, whether they used the facility or not. So this is potentially another hunk of out-of-line hardware for users with special needs. We’ve identified several others of the sort: large look-up tables, Galois Field operations, etc. They would all use the same interface.