What is interesting is that Atom has a higher Stream Memory bandwidth performance but a lower Memory Memory performance. I had to look.
All the memory benchmarks are the same except two std libc routines.
stdlib Write (memset) : iPad 3260, Atom Z2760 654
stdlib Copy (memcpy): iPad 2670, Atom Z2760 1017
This looks like Apple has written custom versions of memset and memcpy like Intel has done with the Intel compiler. If so, the comparison is with a hand tuned ARM libc routine and a dynamically linked Win8 version.
This does not compare silicon performance. It compares the performance MSFT libc routine.
32-bit memory bandwidth on both is similar.
geekbench problem. Insufficient disclosure when comparing two different results.