If you can add FLOPs, and you have a number of cores with the same fixed number of FLOPs/cycle, it follows that you can add their frequency. For the vast major

If you can add FLOPs, and you have a number of cores with the same fixed number of FLOPs/cycle, it follows that you can add their frequency. For the vast majority of real-life workloads there are vast differences between one core having X GHz or Y FLOPs vs lots of cores that sum up to that number. Which is the point GP tried to make.