Accelerated computing continues to gain momentum as the HPC community moves towards Exascale. Our Tesla P100 GPU review shows how these accelerators are opening up new worlds of performance vs. traditional CPU-based systems and even vs. NVIDIA’s previous K80 GPU product. Here are some examples….
These case studies are taken from the insideHPC Research Report on GPU Accelerators.
Caffe/AlexNet: These are deep learning applications that neural networks to recognize images and solve computer vision problems. These apps are widely used throughout the deep learning community and results are often used to compare system performance on deep learning tasks.
As can be seen in the chart, a traditional CPU-only dual-Xeon system is set up with a baseline score of ‘1’ on the comparison. A system equipped with dual K80 GPUs achieved a 5x speed up over the Xeon-only system, and a system equipped with dual P100 GPUs managed a 35x speed-up vs. the Xeon machine.
HOOMD-blue: is a general purpose particle simulation toolkit that allows users to simulate particle reactions under a very wide variety of conditions. HOOMD-blue, along with other particle simulation routines, is an application that greatly benefits from GPU acceleration.
Dual K80 GPUs yield about a 7.5x speed-up vs. a dual Xeon system. Substituting dual P100 GPUs more than doubles the K80 speed-up, and going to 8 P100 GPUs provides more than 30x performance as the base case Xeon result.
MILC: is a set of code that is designed to study quantum chromodynamics (QCD), which is the theory of strong interactions of subatomic physics.
Once again, we see that NVIDIA K80s accelerate this application about 5x when compared with a dual Xeon only system.
Dual P100s drive performance up to around 7.5x vs. Xeon, and using 4 P100 GPUs pushes performance up to 15x. We also again see nearly linear GPU scalability when moving from 2 to 4 to 8 P100 GPUs, with the 8 GPU test coming in with around a 27.5x overall speed-up.
Tesla P100 GPU Review Summary
With more than 400 applications optimized for GPU acceleration – and more coming out every day, we’re fast approaching a time when GPU accelerators will be available for almost every compute intensive workload in HPC and in the enterprise.
Nearly every organization has at least an application or two that can be accelerated through use of GPUs. While higher performance is certainly one of the leading benefits from acceleration, GPUs can provide this performance while using less energy and floor space than traditional servers – which makes them a great choice for enterprises small and large.
See the results of insideHPC readers who are using and testing GPUs. Download the insideHPC Research Report on GPUs