Today's research requires infrastructure that can handle large computational workloads to derive fast and accurate insights from vast amounts of data. Researchers are using OSS and NVIDIA platforms to expand breakthroughs in HPC, AI, and machine learning and data science in disciplines like robotics, autonomous vehicles, and healthcare.
Molecular Dynamics (MD) is a field of research in which a computer simulates the physical movements of atoms and molecules. Molecular dynamics relies heavily on computational power. With accelerated computing, researchers can virtually model millions of molecules at a time, reducing costs and speeding time to solution. The details of the simulation have to be chosen carefully so that the calculation can finish within a reasonable time period while still being long enough for the results to be useful. In order for the conclusions to be statistically valid, the simulated time span should match the kinetics of the natural process. For example, you can’t draw conclusions about how humans walk by looking at only one step. For many applications of Molecular Dynamics, several CPU-days or even CPU-years are needed to process the simulations. Programs that run algorithms in a parallel manner allow the computations to be distributed among CPUs, allowing for more complicated, time consuming simulations.
However, a system with multiple CPUs running parallel algorithms is not the most powerful way to compute molecular dynamics simulations. GPUs such as NVIDIA® A100 Tensor Core GPUs are much faster at processing parallel tasks and thus would be a better option for a molecular dynamics system. The One Stop Systems GPU Accelerators can accommodate anywhere from one to 16 high end GPUs that can be accessed by up to four systems. With this much compute power, simulations can be much longer without taking unreasonable amounts of time; in fact, this much power might allow for simulations that previously would not have been possible.
Here are some molecular dynamics applications that can be sped up by using GPUs:
Numerical analysis is the study of algorithms that use numerical approximation for the problems of mathematical analysis. Exact answers are often impossible to find in math; thus, numerical analysis does not pursue exact answers. Rather, numerical analysis seeks approximate answers that do not cause too many errors. Numerical analysis applies to many different fields including engineering and the physical sciences, such as astronomy, medicine and biology. More recently, it has also applied to life sciences and even the arts. Before computers, numerical analysis was done by hand; now, computers are used to calculate instead. There are many widely used numerical computing applications such as MATLAB, ArrayFire, Mathematica Wolfram, and NMath Premium that can benefit from the additional compute power GPUs can provide. The more GPUs used, the faster the algorithms in numerical analysis can be computed.
In addition to fast computation, numerical analysis requires a large amount of storage because of the sheer amount of data required. There are many errors that can arise in numerical analysis. Studying these errors is an important part of the field. For example, round-off errors occur because it’s not possible to represent all real numbers on a system that has finite memory. However, more memory can make a difference. A typical server in the high-performance computing world may be 4U and have multiple internal terabyte drives. If you connect this one server to the 3U OSS Flash Storage Array, it would then have over 200TB of memory. In order to achieve this amount of storage with servers alone, it would take over 780 of the 4U, 256GB servers (over 3000U of rack space). Instead, it can be done in 7U rack space. Imagine the benefit this much flash storage could add to a field like numerical analysis.
Applications such as ocean circulation modeling, tsunami simulation, ocean modeling, computational fluid dynamics and weather and climate forecasting rely on technology to be as accurate as possible. For example, Weather and climate forecasting uses mathematical models of the atmosphere and oceans to predict the weather based on current weather conditions. Once computers were invented, realistic results were finally possible. And as technology has advanced, the forecasts have become more accurate. Global and local forecast models are run all over the world and they use current weather observations received from different sources, such as weather satellites. Utilizing OSS platforms with the help from NVIDIA, deep learning training and inference is transforming the way computers perform perceptual tasks such as computer vision, pattern detection, speech recognition, and behavior prediction. These improvements make it possible to take labor-intensive jobs and automate them for applications like disaster response and humanitarian relief.
Weather and climate forecasting can be used to create short term weather forecasts, or long-term climate forecasts. This has been helpful for interpreting and predicting climate change. In regional models, improvement has been made when tracking tropical cyclones and predicting air quality. All weather and climate forecasting models are made up of huge amounts of data. Computing the vast amount of data and executing the intricate calculations required for modern weather and climate forecasting necessitates the use of some of the most powerful supercomputers in the world.
NASA Goddard Space Flight Center's High-End Computer Network (HECN) Team is utilizing two of our 2U Compute accelerators to demonstrate high speed disk-to-disk transfers. Each enclosure supports up to eight RAID controllers that transfer data to solid-state disk drives, achieving disk-to-disk transfer speeds that exceed 100Gb/s. These unprecedented data transfer speeds allow NASA to predict climate change and other complex modeling and simulation tasks.
Here are some weather and climate forecasting applications that can be sped up by using GPUs:
Our compute accelerators support from one to sixteen double-wide PCIe cards and can be cabled up to four host computers through PCIe x16 Gen3 connections each operating at 128Gb/s. The all-steel construction chassis house power supplies, fans, and a system monitor that monitors the fans, temperature sensors and power voltages. Front panel LEDs signal minor, major or critical alarms. The compute accelerators are transparent and do not require software except for the drivers required by the PCIe add-in cards. Compute accelerators are the best appliance for applications that require a large amount of compute power.Learn More
Our flash storage arrays support from one to thirty-two single-wide PCIe cards and can be cabled up to four host computers through PCIe x16 Gen3 connections each operating at 128Gb/s. The light-weight, compact chassis house power supplies, fans, and a system monitor that monitors the fans, temperature sensors and power voltages. Flash storage arrays are the best appliance for applications that require a large amount of fast, flexible storage.Learn More