If your calculation is correct, how come you're using 160 MB on 32 bit C++14? The reason why you're getting MLE is because new has a large memory overhead. You can try it yourself with custom invocation.

The code below uses almost 32 MB on custom invocation, even though only 16 MB was requested. To get rid of this overhead, allocate a large memory buffer and then write your own replacement of new which increments a counter and returns a new address each time. This idea is called Arena allocation.

using ull = unsigned long long;

struct _16bytes {
    _16bytes* a[2];
};

int main() {
    ull q;
    for (int i=0; i<1000000; i++) {
        q ^= reinterpret_cast<ull>(new _16bytes());
    }
    return q == 0;
}