I believe that’s inaccurate, at least on a modern CPU. The bookkeeping for the stack is faster, since ‘allocating’ and ‘deallocating’ is just subtracting from and adding to a register. And the area of the stack in active use at any given time is usually tiny (well under a kilobyte, even though the full stack is usually several megabytes), so it’s likely to stick around in L1 cache. And return addresses on the stack get special treatment by the branch predictor. But other than that, it’s treated the same as any other memory.