SimBioSys Blog

We’ve been having the conversation within our company that the two dials of speed and accuracy work counter to each other. So, we’ve been espousing that even when it comes to the eHiTS Lightning solution that higher accuracy does take longer. We still stand by that BUT what we are happy about is the type of accuracy we can achieve very quickly using the new eHiTS Lightning algorithms. This becomes more obvious when our results are compared to the results of others. There has been a proliferation of arguments for GPUs being used as acceleration processors – we actually believe this is simply because of the business driver of “looking for new markets” for the GPU manufacturers. Zsolt has discussed his views regarding the future of High Performance Computing previously and commented on GPUs. Our belief is that while GPUs are clearly more “common” our decision to work with the Cell BE processor can certainly lead to far superior results…don’t forget that the RoadRunner computer is based on the Cell Processor, not GPUs. Did we make the right decision?

We are always watching for innovative solutions in docking. We acknowledge those scientists pushing towards the edge of performance and excellence. When we saw the recent announcement regarding the DockStar solution from Silicon Informatics we were interested to see whether they had made some of the promised breakthroughs with their GPU-based solution. Their website promises “With the combined power of the DockStar™ Linux Workstation, NVIDIA’s® Tesla™ GPU’s and our proprietary software kernels, Silicon Informatics’ DockStar™ solution outperforms conventional workstations by 10 - 20+ times.” The system is based on the Autodock 4.0 software platform. As commented in my recent blogpost we have been doing a lot of work to validate the performance of eHiTS Lightning and gathering validation data for throughput, pose accuracy and enrichment so we were interested to compare our data with those of the GPU-based DockStar solution. We’ll report the data in much more detail in a Case Study note presently in development but our observations at present are based on comparing to information they have on the site.

There are 3 examples posted on the home page of the DockStar site, 1stp, 3ptb and 1hvr, with the results shown below:

The table shows us that for these three examples at least we see a difference of over 10x in performance for the Cell processor versus the GPU-based Dockstar solution. Now, this is only a comparison based on speed. Accuracy is clearly just as important so how do we do there?

We are presently finishing the results for all examples but one example is shown below, in all its glory! Notice the dramatic performance difference in the plots below. The eHiTS Lightning shows the expected behavior in terms of the expected good, i.e. low scores at low RMSD values whereas DockStar/AutoDock accuracy / score correlation has no tendency. These results show that eHiTS Lightning not only offers dramatic speed advantages but also the accuracy advantages we have been espousing. More detail will be published soon.

	Img1: Autodock 4: 250,000 GA: 45 minutes, note the resultant RMSD distribution.
	Img2: eHiTS Lightning, on the CELL B/E. 1 minute, note the nature of the Scrore/RMSD distribution, most poses are at low RMSD values.