Fast sort on CPUs and GPUs: a case for bandwidth oblivious SIMD sort