P C T A L K HIGH SPEED SOLUTION The switch to GPUs for high-performance computing offers a competitive advantage for those willing to invest in the technology. By Charles Clarke There is a quiet revolution taking place in the world of CAE. With the introduction of new graphics processors from Nvidia and AMD/ATI, CAE developers can now off-load some of the more compute-intensive parts of the CAE simulation calculations to these massively parallel graphics accelerators, slashing CAE solution times by up to 80% of conventional multi-core CPUs. COMPARING CPUs with GPUs Central processing units (CPUs) are highly versatile processors with large, complex cores capable of executing all routines in an application. They are used in the majority of servers and desktop systems. Compared with CPUs, graphics processing units (GPUs) are more focused processors with smaller, simpler cores and limited support for I/O devices. Recent generations of GPUs have specialised in the execution of the compute-intensive portions of applications. They are particularly well suited for applications with large data sets. Application development environments for GPUs use techniques that allow the GPU to handle compute-intensive portions of applications that usually run on CPUs. The solution of the simulation equations is relatively trivial, in comparison to the complexities of their formulation and yet complex analyses have been constrained by hardware capability for decades. As faster hardware became available with corresponding increases in memory capacity and speed, so analyses got ever more complex. Today Multi-physics simulations are commonplace and Automotive designers can do reliable CFD simulations of whole cars instead of just the discrete components they were analysing just a few years ago. These complex analyses still take significant time to solve with solution times being measured in hours and sometimes days or even weeks. Quite obviously reducing solution time is the Holy Grail, not just to reduce overall project development times, but when the solution time is shorter, you find out you’ve made a mistake sooner and similarly contribute to reduced project times. Time was when any kind of computationally intensive activity involved large air-conditioned rooms full of very expensive super cooled equipment – the stuff of science fiction movies. The kinds of problems that these high-powered computers were directed at, were geo-physical problems like climate modelling or seismic analysis, in fact any kind of ‘domain’ problem that required the solution of millions of simultaneous equations. The mathematics is not complicated – just enormous and highly iterative (hence the use of these number crunchers in the first place). Complex multi-physics finite element modelling and CFD simulation fall into this category. FASTER PROCESSING Fundamentally, there are only two ways to speed up a computer. You can increase the system clock frequency by improvements in silicon technology, thereby processing information faster, or you can attempt to do more processing in each cycle, effectively doing work in parallel. During the late 1970s and up to about 1985 there was about a 30% performance gain every year because of improvements in semiconductor technology alone. From 1985 to the turn of the century the computer industry has recorded a 700% performance gain from improvements in semiconductor logic and the exploitation of all available forms of parallelism. There are basically three levels of parallelism, uniprocessor parallelism, multiprocessor parallelism and multicomputer parallelism. In the quest for more and more performance there is no substitute for very fast single processors. The better the performance of the processor the greater the overall system through-put. Similarly, parallel systems made from fast processors, whether singly or in clusters, benefit from these higher levels of performance. However, the ultimate in processor performance is very expensive. It needs very fast memory, water cooled, Gallium Arsenide chip technology, very specialised operating systems and system administration software. These are characteristics found in the ‘old’ multi-million dollar Cray computers. To achieve a more affordable alternative to Cray, the performance gains arising from 40 July 2017
EN-July2017-eMag6
To see the actual publication please follow the link above