ReviewInformaticsThe impact of accelerator processors for high-throughput molecular modeling and simulation
Section snippets
Accelerator processors
Historically, microprocessor performance has improved primarily through the raising of clock speeds, make possible by the development of ever-finer fabrication processes. In recent years, it has become increasingly difficult to keep increasing clock speeds because of fundamental limits in the process technology and power consumption. Performance has also been limited by the increasing relative cost of accessing main memory, the speed of which has increased at a slower rate than CPUs. Despite
Software re-engineering
Although APs offer high peak computational performance, exploiting this efficiently comes at the cost of ease of use: codes must be refactored as highly parallelized programs. Code redesign and redevelopment has a very high cost that, when weighted against the traditional ‘free’ performance increase provided by the next iteration of conventional hardware, has in the past significantly limited the appeal of special-purpose hardware on the high-performance computing market. In the next few years,
Accelerated modeling
Despite the very recent introduction of APs into the market, a variety of applications targeting different industrial and scientific fields have already appeared. Excellent performance speed-ups have been reported for such diverse cases as computational finance [22], fluid dynamics [23], sequence alignment [24] and quantum chemistry [25]. These notable achievements for such a variety of algorithms highlight the potential of APs for computational science in general.
Accelerators processors will
Distributed computing in the accelerated era
Computational grids enable scientists to distribute simulations across a pool of machines to benefit from their aggregate power, and have already proved useful for fast calculations of binding affinities using MD techniques [31]. Owing to the computational cost of molecular simulations, the grid is usually composed of very expensive parallel-processing HPC resources, the costs of which limit the applicability of the approach. When distributed computing is combined with AP-equipped hardware,
Future outlook for medium-throughput molecular modeling
There is great interest in methods for supporting and optimizing experimental high-throughput screening 4, 38 to identify, characterize and optimize possible leads for a given target out of the vast number of viable chemical compounds. It is, however, very difficult for such methods to account correctly for the many phenomena involved in complex formation, such as the subtle interplay between entropy and enthalpy, conformational changes of the ligand or the substrate, presence of water
Conclusion
APs have the potential to provide a radical change in scientific and industrial computation. With an effective performance tens of times that of standard computers and doubling each 8–12 months, unprecedented levels of computational power can be put into the hands of scientists and programmers at an affordable cost. Such disruptive technology is already being used by research groups and companies in many different fields, and applications targeting molecular modeling are appearing at a steady
Conflicts of interest statement
We notify the journal that the authors are scientific consultants and also share holders of Acellera Ltd, a UK-based company selling software solutions for accelerated processors.
Acknowledgements
We gratefully acknowledge support from Barcelona supercomputing center (http://www.bsc.es), Acellera Ltd (http://www.acellera.com), Sony Computer Entertainment Spain (http://www.scee.com) and Nvidia corporation (http://www.nvidia.com). We thank Jordi Mestres and Ferran Sanz for a critical reading of the manuscript. GG acknowledges the Aneurist project (http://www.aneurist.org) for financial support. GDF acknowledges support from the Ramon y Cajal scheme.
References (44)
- et al.
The myth of the biotech revolution
Trends Biotechnol.
(2004) Hardware accelerator for molecular dynamics: MDGRAPE-2
Comput. Phys. Commun.
(2003)Performance of the cell processor for biomolecular simulations
Comp. Phys. Commun.
(2007)- et al.
Reengineering the pharmaceutical industry by crash-testing molecules
Drug Discov. Today
(2005) VMD: visual molecular dynamics
J. Mol. Graph.
(1996)Just add water! the effect of water on the specificity of protein–ligand binding sites and its potential application to drug design
Chem. Biol.
(1996)- et al.
Key factors in the rising cost of new drug discovery and development
Nat. Rev. Drug Discov.
(2004) The many roles of computation in drug discovery
Science
(2004)In silico pharmacology for drug discovery: methods for virtual ligand screening and profiling
Br. J. Pharmacol.
(2007)- et al.
Understanding Molecular Simulation
(2002)
Lithography and the future of Moore's law
Proc. SPIE
Scalable molecular dynamics with NAMD
J. Comput. Chem.
Anton, a special-purpose machine for molecular dynamics simulation
Cited by (32)
pointerchain: Tracing pointers to their roots – A case study in molecular dynamics simulations
2019, Parallel ComputingCitation Excerpt :It simulates the physical movements of atoms and molecules with a Hamiltonian of N-body interactions. Over the past three decades, we have witnessed the evolution of MD simulations as a computational microscope that has provided a unique framework for understanding the molecular underpinning of cellular biology [4], which applies to a large number of real-world examples [5–11]. Currently, major MD packages, such as AMBER [12], LAMMPS [13], GROMACS [14], and NAMD [15], use low-level approaches, like CUDA [16] and OpenCL [17], to utilize GPUs to their benefits for both code execution and data transfer.
Crowd computing: Using competitive dynamics to develop and refine highly predictive models
2013, Drug Discovery TodayCitation Excerpt :Through the BOINC platform [5] it is now straightforward to parcel a problem onto a distributed network of volunteer computers; participants donating computational resources need to have but a fleeting interest in the underlying scientific problem. With the inclusion of advanced graphical processing units (GPUs) and internet accessibility as standard for seventh-generation gaming consoles, scientists began to use these devices as a way to harness the sheer number, and commensurate power, of consumer-focused gaming machines [6,7]. Gamers are engaged somewhat passively – they play the games, the scientists perform ‘the science’ in the background.
High-throughput molecular dynamics: The powerful new tool for drug discovery
2012, Drug Discovery TodayBubbles and solid structures in a vibrated bed of granular materials
2011, Physica A: Statistical Mechanics and its ApplicationsImplementing molecular dynamics on hybrid high performance computers - Short range forces
2011, Computer Physics CommunicationsCitation Excerpt :Another complication that arises in software development for hybrid machines is the fact that the instruction sets differ for the CPU and accelerators. While software developers have had the benefit of “free” performance gains from continuously improving x86 performance, software for hybrid machines currently must be ported to a compiler suitable for the new architectures [8]. It is undesirable to port an entire legacy code for use on accelerators [22], and therefore minimizing the routines that must be ported for efficient acceleration is an important concern.
GPU-accelerated molecular modeling coming of age
2010, Journal of Molecular Graphics and ModellingCitation Excerpt :There has been a great deal of interest in the use of accelerator devices to augment multi-core CPUs for computationally demanding molecular modeling applications as a result of the relentless drive for increased performance, improved price/performance, and greater performance/watt efficiency. Previous efforts with accelerators, including field-programmable gate arrays (FPGAs) [11,12], several generations of MDGRAPE [13–15], the Cell processor [9,16–20], and GPUs, have demonstrated the potential performance benefits available to molecular modeling, while bringing to light many of the software engineering challenges involved in adapting existing applications for heterogeneous computing. Although the specific details of programming GPUs differ somewhat from other accelerators, all heterogeneous computing approaches involve porting or adapting existing algorithms for the target accelerator device, managing multiple independent memory spaces, balancing work among host CPUs and accelerator devices, and coping with host-device communication latency and bandwidth limitations.