Exascale Computing. We previously mentioned Xxxxx'x law. Xxxxx’x law states that the number of transistors in a dense integrated circuit doubles approximately every two years. This law formulated in 1965 not only proved to be true, but it also translated into the doubling of the computing capacity of cores every 24 months. This was possible not only by increasing the number of transistors but also by increasing the frequency at which they worked. However, we can say that in the last few years the end of this paradigm seems to have arrived. The reason is that the performance of a single core is no longer increasing at the same pace. There are three main reasons for this: The memory wall: This refers to the gap in performance that exists between processor and memory (CPU speed improved at an annual rate of 55% up to 2000, while memory speed only improved at 10%). The main method for bridging the gap has been to use caches between the processor and the main memory, increasing the sizes of these caches and adding more levels of caching. But memory bandwidth is still a problem that has not been solved. The instruction-level parallelism wall: Increasing the number of transistors in a chip as Xxxxx’x law says is used in some cases to increase the number of functional units, allowing a higher level of instruction-level parallelism (ILP) because there are more specialised units or several units of the same kind (i.e., two floating point units that can process two floating point operations in parallel). However, finding enough parallelism in a single instruction stream to keep a high-performance single-core processor busy is becoming more and more complex. One of the techniques to overcome this has been hyper-threading. This involves making a single physical core presented as two (or more) to the operating system. Running two threads allows exploitation of instruction-level parallelism. The power wall: As we stated above, not only the number of transistors has been increasing but also their frequency. Yet there exists a technological limit to surface power density, and for this reason, clock frequency cannot scale up freely any more. Not only would the amount of power that must be supplied be unfeasible, but also the chip would not be able to dissipate the amount of heat generated. To address this issue, the trend is to develop simpler and specialised hardware and aggregate more of them (i.e., Xeon Phi, GPUs). Nevertheless, computer technologists are also focussed on making exascale ma...
