If I want a 1000x1000 array, representing it physically as a single 1000000-element array requires one allocation, and processing it element-by-element (assuming it's stored in the same order we're iterating over it) is sequential in memory and therefore very efficient.
Representing it as 1000 separate 1000-element arrays requires 1000 allocations, and pointer-chasing every time we move from one row to the next.
Otherwise you would have an array of pointers to arrays. The usage (syntax) for them would be the same but the performance would not be.
They also have different uses. You would expect an array of arrays to be an array of arrays which share the same length. For an array of pointers to an array you would expect dynamic length arrays contained within the original array.
Even in c++ could you not just define some int [1000][1000]foo? I've never really used C++ but my C knowledge assumption is that is 1000000 continuous elements.
std::array<std::array<T, N>, M> data;
int data[M][N];
Edit:
For the stack at least. On the heap, you'd need to use a single std::vector<int> and do the indices manually, or use mdspan
It works fine in C though, or FORTRAN, or Ada, or ALGOL 60, ...
NVidia has pivoted to design CUDA hardware with focus on C++ back in , and seems to be doing quite well for them.
CppCon 2017: "Designing (New) C++ Hardware”
https://www.youtube.com/watch?v=86seb-iZCnI
They were also the ones sponsoring the ISO work on mdspan, while HPC research labs are pushing for linalg on top.
I would rather be using Ada today, but that isn't how the world moves.
If it fits on the stack, yes.
Typical code using MD-arrays is scientific code, and the data they manipulate generally do not fit there.