As the law of Moore decelerates CPUs, committed graphics co-processors are capturing some of the slack. Just like GPUs are altering the game in intense learning and autonomous cars, the desktop PC powered with GPU might even instigate to maintain the pace with the traditional supercomputer for a part of supercomputer applications.
For example, a team of scientists from Russia is reporting this month that they have been successful to solve computational issues in nuclear physics with the use of an off-the-shelf, powerful PC incorporating a GPU. According to them, after fine-tuning the algorithms for GPUs, it was possible for them to run their calculations much swifter than the conventional, CPU-powered supercomputers that were used by others.
“I have no queries regarding that many other researchers across the globe could reach to similar results in their domains like seismology, geophysics, medical diagnostics, and more,” says Vladimir Kukulin, professor of theoretical physics at Lomonosov Moscow State University. “But only by linking two main ingredients – Reformulation of the entire issue and then by inventing some effectual method on how to parallelize the entire execution in multiples of even millions of free threads,” says Vladimir.
The major issue that the Kukulin’s team was encountering involved the wide calculations required to brief scattering problems in their domain – like when a single nucleon collides with a particle or any other nucleon generates a spray of particles and results in daughter nuclei. Such nuclear numerous-body trouble, Kukulin says, could need calculations involving matrices comprising millions of particles.
Matrix Algebra with many moving parts can impasse a supercomputer. But, according to Kukulin, his team realized too, that calculations with huge matrices could intend independent threads of instructions to perform in parallel with numerous other similar threads. The parallelizability of his team’s nuclear calculations, in other terms, implies that it was a prime candidate for operating effectively on a GPU.
The GPU, basically introduced to handle matrix-bulky calculations required to generate real-time graphics, is discovering unexpected applications in some domains today, including molecular modelling, Bitcoin mining, and other similar applications. Kukulin says that the list of computational movements that the GPU can manage is a list that now contains nuclear physics, will only expand.
As a whole, he confirms that such problems that could borrow themselves to GPU supercomputing on the cost-effective are those who individual materials are not interlinked with each other. It is because interdependence implies that individual elements would have to pass through a regular logic passage, checking for each and every element’s current influence on other elements involved in the calculation.
“You must write your issues into a form that enables you to collectively parallelize,” he says. “It is essential to avoid by some means any conditional.”
According to Kukulin, 3D ultrasound imaging, a computer-based medical diagnostic division could be more extensively embraces if medical units require only a couple of thousand dollars for a desktop PC as against to numerous thousands of dollars for simple access to supercomputers. In the end, the troubles fit for speeding up the GPU need not only programming but also efficiency in the domain of application to figure out the precise way to such issues.