Today, the main problem faced by researchers who use computing systems or HPC, are:
- Systems consume more and more. The lower cost of hardware has produced that there are more and more clusters with more nodes and consumption is becoming worrying. In addition, to the heat factor that they generate, we must add a factor of between 1,3-1,5 (depending on the efficiency) of consumption of the air conditioning systems.
- Space is very limited. The CPDs in which they are housed no longer have room in the racks and there is also no excess in the refrigerated rooms of the departments.
- If we want to parallelize the applications between nodes, it is necessary to implement Infiniband infrastructure, Switch and other elements that suppose an extra cost.
- The larger the cluster, the greater the complexity in its administration and management, which makes its cost of ownership more expensive.
- Researchers using commercial programs must pay a license per machine, which in a cluster shoots up ownership costs and annual software maintenance.
WHY MIGRATE FROM CPU TO GPU
The inconvenience of migrating to GPU involves rewriting all the code. This is a problem with departments that have their own Fortran applications or old versions of commercial applications.
However, many departments use basic development tools like Matematica and Mathlabs or computer vision libraries like OpenCV. In these cases, everything developed on this basis can be directly migrated to GPU, since these applications already support it themselves. In addition, young researchers create their new codes from scratch, which allows them to start developing on GPUs, which will be the majority reality on the market in a short time.
The problem in terms of programming difficulty has been largely reduced, thanks to NVDIA and its CUDA technology. The compatibility from version 1.0 to the current 5.5 and the imminent release of 6.0 provide greater simplicity, but above all a guarantee regarding the work done for the future.
As we can see in the graph, the leading applications in sectors such as molecular dynamics, material sciences, earth sciences, fluid dynamics and other branches of physics, achieve GPU performance of between 10 and 20 times on average, higher than the obtained with high-performance CPUs, as we can see in the graph.
This means that the performance improvement is spectacular, in some applications like NAMD or AMBER. In these applications, we can be talking about 8 to 10 times faster performance for a GPU compared to a current CPU. This means from a practical point of view, that one node with GPUs can easily equal 10 or more nodes that your department has so far.
Currently, more than 70 benchmark applications add support for graphics card accelerators (GPUs) to meet the demand for faster simulations. Application developers embrace accelerated computing, which will allow users to design higher-quality products and deepen their scientific knowledge.
In this line, SIE offers solutions with GPU from Workstation for only 2.000 euros and low noise, to be able to comfortably place entire clusters of GPUs connected by Infiniband solutions in your office.
In fact, currently a large part of the TOP500 clusters are based on NVDIA GPU solutions, with CUDA programming. This new technology allows thousands of Tflops/s to be offered with half the consumption of CPU-based solutions.
Finally, we would like to mention a series of applications, most of them tested in SIE, and that work perfectly with our SIE Ladón GPU equipment.
All of them already support multi-GPU technology in their latest versions, which means that our systems can currently integrate up to 4 GPUs and in the immediate future up to a total of 8 GPUs and take advantage of the capacity they offer.
The most important applications, part of them Spanish are:
- MOLECULAR DYNAMICS
- ACEMD, AMBER, DL-POLY, GROMACS, LAMMPS and NAMD
- QUANTUM CHEMISTRY
- Abinit, GAMESS-UK, GAMESS-US, MOLCAS, NWChem, Octopus and TeraChem
- NUMERICAL ANALYSIS
- Mathematica Wolfram and MATLAB by Mathworks
COMPUTATIONAL FLUID DYNAMICS AND STRUCTURAL MECHANICS
ANSYS Fluent, Ansys Mechanical, OpenFOAM (FluiDyna Culises) and Abaqus
With these applications, we have great references from our clients, without forgetting that SIE sells many of them.
Next, we will expand a little information about these technologies.
WHAT IS THE GPU (GRAPHIC PROCESSOR UNIT) COMPUTING
"GPU computing first saw success with researchers who could use CUDA to accelerate their own applications in their scientific research and discovery," said Addison Snell, CEO of Intersect360 Research. “We have now entered a new era of more commercial GPU-optimized software, expanding acceleration opportunities across a range of enterprise and engineering computing solutions.”
This is a partial list of other GPU-accelerated applications released or in development:
- Computer Aided Engineering: Abaqus/Standard, Agilent ADS and EMPro, ANSYS Mechanical, CST MWS, MSC Nastran, Marc, OpenFOAM Libraries, RADIOSS™
- Defense and intelligence: DigitalGlobe Advanced Ortho Series, Exelis (ITT) ENVI, Incogna GIS, Intergraph Motion Video Analyst, MotionDSP Ikena ISR, PCI GEomatics GXL
- Media and Entertainment: Adobe CS6, Autodesk 3ds Max and Maya, Blackmagic DaVinci Resolve, Chaos V-Ray RT, Elemental Server, Telestream Vantage
- Oil and Gas: Acceleware AxRTM, ffA SVI Pro, Headwave Suite, Paradigm Echos RTM, Schlumberger Visage, WesternGeco Omega2 RTM
- Scientific calculation: AMBER, CHARMM, Chroma, FastROCS, GAMESS, GROMACS, GTC, WL-LSMS, MATLAB, MILC, NAMD, QUDA, VASP, VMD
- Weather and climate models: COSMO, GEOS-5, HOMME, HYCOM, WRF, NEMO, NIM
- The full list is at http://www.nvidia.es/object/gpu-computing-applications-es.html
The most accessible parallel processors
The advent of computational accelerators with powerful parallel processing architecture, easily programmable in high-level and widely used languages, or employing self-parallelizing compilers, has made it possible for developers to maximize application performance.
Accelerators provide the developer with a high degree of flexibility to take advantage of the drastic increase in speed of applications with languages as widespread as C, C++ and Fortran or the programming model based on the directives of the OpenACC standard.
The basic extension of these high-level programming languages allows parallelism to be introduced with the NVIDIA CUDA parallel computing platform and programming model. The CUDA platform can now be used with all NVIDIA GPUs, giving it a worldwide installed base of more than 415 million CUDA-enabled GPUs.
What is CUDA
It is a parallel computing platform and programming model developed by NVIDIA. It harnesses the enormous power of GPUs to provide an extraordinary increase in system performance. It is currently in version 5.5 and version 6 is about to come out and it is compatible from the first versions 1.0. This allows all NVDIA cards to support the CUDA standard.
NVIDIA
NVIDIA showed the world the possibilities of the graphics chip with the invention of the GPU in 1999. Today, its processors are the foundation of a wide variety of products ranging from smart phones to supercomputers. NVIDIA mobile processors are used in mobile phones and tablets, and in vehicle infotainment systems. PC gamers use GPUs to bring spectacular worlds to life. Professionals use them to create visual effects in movies and make everything from golf clubs to large commercial airplanes. Finally, researchers harness the power of the GPU to drive the advancement of science through high computing systems. The company owns more than 5000 patents worldwide, some of which provide essential designs and concepts for today's computing. For more information, visit www.nvidia.es.