The list of Top 500 supercomputers is incorrect

The problem lies in that the way of evaluate the performance only measures the quantity of petaflops (floating point operations by second). For this they use a benchmark software called Linpack, that returns the number of petaflops measuring the time that spend the supercomputer executing a lot of algebraic operations. The problem is that the supercomputers don't only make algebraic operations. Also are importants, the speed with that can store and load data from the memory, instructions for the control flow of the program, etc.

That a supercomputer can execute a lot of operations in floating point by second is good to measure how will be the performance in applications that make a lot of calculations of this kind, like in 3D physics simulations, but this is not the only type of application thought for supercomputers. Since that the supercomputers are not used for a unique scientific project and after are throw away, but usually works with different kinds of applications, would be convenient don't fabricate them thinking in only a kind of application.

Now they are taking into account of this situation and have emerge a new initiative called Graph 500, where the word "Graph" refers to a mathematic tool, that is a graphic representation of the elements of a set. The intention is change the way of evaluate the performance, using differents benchmarks focused in applications with an intensive use of data, for instance, in medical applications, biology, data mining, social networks, etc. In this way, the hardware engineers will know better how focus their designs at the time of try to get the maximun performance for this kind of applications. The benchmarks selected execute algorithms that work with graph. These are: concurrent search, single source shortest path and the maximal independent set.

Therefore now we have on the one hand, the Top 500, that measure the calculus power of the machine and, by the other hand, the Graph 500 that measure the capacity to work with high volumes of data. In this way the people that go to rent a supercomputer will can choose better, according to their needs, consulting these lists. Although the ideal way of measure the performance always is measuring how long it takes the application that we need execute in that machine, but that isn't always possible in supercomputation.