March 16, 2010 | Guest Commentary | Several years ago, Brian Beck and I founded the Nevada Center for Bioinformatics in Reno. We got along great, but when it came time to buy computing systems, we were at the opposite end of the spectrum. I wanted to get an FPGA accelerated system, but it did not meet his purposes. He wanted to get a standard Linux cluster to run his molecular modeling code, whereas I was more interested in sequence analysis. We wound up getting a slightly smaller cluster and a single accelerator node. This made us both happy, as I was able to run enormous Hidden Markov Model (HMM) searches on the accelerator without monopolizing the cluster for weeks on end.
Today, the choice might have been different. The Tesla Bio Workbench was recently announced by Nvidia, maker of Graphics Processing Units (GPUs) for the computing industry and lately for high-performance computing (HPC) as well. Nvidia processors were originally designed for graphics processing, but have been adapted to general-purpose computing through the use of a computing architecture known as CUDA. Bio Workbench is simply a website (www.nvidia.com/object/tesla_bio_workbench.html) that brings together all of the algorithms for the life sciences for the convenience of the researcher.
The number of algorithms that are available to run on GPUs is rather impressive. On the sequence analysis side, BLASTP, HMMer, Smith-Waterman, MUMmerGPU, ClustalW, MEME, and Infernal are all available for download. For those interested in docking, algorithms such as autodock and piper have shown impressive speedups of 10 to 16 times in tests.
Molecular Dynamics fans will find plenty to like at the bioworkbench website, with AMBER, GROMACS, HOOMD, LAMMPS and NAMD all demonstrating significant speedups. VMD is available to animate and analyze large biomolecular systems at up to 100 times faster than on a standard CPU. TeraChem is a general-purpose quantum chemistry package that has been shown to demonstrate as much as a 50 times speedup.
It is wonderful that these programs are available for the price of a download! Perhaps the best part about accelerating with GPUs and CUDA is the remarkably low cost of entry. A $50 graphics card is all that you need to get into the game, and scalability is only limited by the size of your pocketbook. The Tesla card and servers hold the high end with what Nvidia claims is the world’s first Teraflop processor.
This amount of power does not come without caveats. Be sure to check the licensing agreements to make sure that you do not violate any copyrights. Nvidia does not provide support or training, and so these programs must be treated as any other open-source software.
Other limitations may exist with these programs. GPU-HMMer, for example, only has accelerated the hmmsearch algorithm, not the hmmpfam program. Also, the latest version of HMMer uses a different format for the models, so you will want to verify compatibility before investing too much time.
Does this mean that GPUs will take the bioinformatics speed crown from accelerators based on FPGAs? I don’t think so. There are still performance advantages whether one chooses a preprogrammed supplier or a ‘do-it-yourself’ solution using commodity cards. Martin Herbordt, for example, found that FPGAs gave 8 times better performance on CHARMM energy minimization calculations than those same calculations on an Nvidia Tesla processor.
Still, the ease with which GPUs can be programmed gives them an important and growing place in the field of HPC and in the area of bioinformatics in particular. If Brian and I were setting up a statewide resource today, we might have been able to find compromise in an accelerated cluster and enjoyed the best of both worlds.
Martin Gollery is founder of Tahoe Informatics. He can be reached at Marty.firstname.lastname@example.org.