Can GPUs Trump Core Count?
When considering the specifications of an engineering workstation, there are certain “musts.” You must consider CPU specs (number, cores, speed). You must consider RAM. And you must consider the graphics card.
But what if there were another “must”? What if a second GPU — one not dedicated to the display — could amplify the performance of all the other components in the workstation? The performance-enhancing GPU is rapidly leaving “what if” behind. It is becoming one of the key components that separate the high-performance engineering and scientific workstation from a gamer’s computer that somehow ended up on an engineer’s desk.
The positive impact of GPUs
To understand why GPUs can have such a positive impact on system performance, the first thing you have to do is get past the notion that GPUs are only good for rendering images at high speed. Sure, they do precisely that, but modern GPUs are also high-performance processors that can be tasked with many different jobs. The reason they can race through them so quickly is parallel processing. Where a high-performance CPU might have 8 cores, a GPU can have over 100. They’re relatively simple cores, but if a set of instructions is properly divvied up and parallelised, the results can be amazing.
Because there are so many cores in a GPU, parallel processing techniques can work even if there is a single GPU assisting the CPU. And because so many engineering and scientific problems lend themselves to parallel solutions, the combination could completely change the level of performance engineers expect from their workstations.
The software will follow
Engineering workstations built around a CPU/GPU combination are only now beginning to appear on the market. As they appear and become more common, so will the software to manage the applications and take advantage of the parallel capabilities.
Right now, the task scheduling software exists, but its use is mainly in supercomputing centres, where GPU-based computing behemoths compete to be the fastest in the world. This scheduling software is almost exclusively Linux-based and tends to be customised for each supercomputer, though there are open-source frameworks available to get teams started. It is inevitable that versions written for GPU-based workstations will appear as more systems appear on the market.
GPUs enter the high-performance engineering workstation world in another way, as well. As the processing power increases, the need to see the result increases as well. Advanced visualisation is as much a part of modern scientific and engineering computation as the basic generation of numbers.
The workstation that this combination of needs requires is increasingly one in which a team of GPUs is managed and coordinated by a CPU sitting in the middle waving a digital conductor’s baton. The product is advanced but not science fiction. If you don’t have this sort of GPU-centric workstation in your one year plan, you should consider it an integral part of your five-year hardware refresh plans.