Sunday, November 6, 2011

Molecular dynamics and supercomputing, past and present

What is the most powerful single computer you can get today, measured in sheer numbers of double precision floating point operations per second (flops)? Possibly, the answer is a rack server (such as the HP half-width 4U in the picture below) with no less than eight(!) Nvidia Tesla M2090 cards installed. Each card offers a theoretical max of 665 Gflops of double precision computing power -- a total of 5.3 Tflops (not counting any CPUs). Now, go to the Top500 and find the list from 10 years ago ... November 2001. Just one of these servers would bring you a safe 3rd place on the list - very impressive for just ONE single machine.

Foto

One might argue, that a 4U rack unit is not a very practical computer - so what's the max you can cram into a a standard sized ATX chassis these days? Well, take four of these GPUs (and a good power supply), and you're up to about 2.7 Tflops, theoretical peak. 10 years ago that would bring well into top10 world wide with a box that you could put under your desk!

The Japanese "K Computer" is currently the fastest cluster out there and recently broke the 10 Pflops barrier (10¹⁶ flops)! Now, he question is, when can I expect to have that kind of performance in a desktop PC, and how about my huge ms protein molecular dynamics simulation?

Speculating on the future is difficult, so I'm just going to "re-blog" this review paper I stumbled upon by Michele Vendruscolo and Christopher M. Dobson, "Protein Dynamics: Moore’s Law in Molecular Biology", Current Biology, 21, R68-R70 (2011), which speculates on the future of MD simulations. According to the authors, MD computing power sees exponential growth, so hopefully, I can one day have a 10 Pflops (I'm willing to settle for less) box under my desk. Here are some of the bold predictions from the paper:


  • In 2010 we saw the first nanosecond scale MD simulation of the Ribosome
  • In 2030 we'll be moving up to millisecond  scale MD simulation of the Ribosome
  • in 2050 we'll be doing all-atom dynamics of entire bacteria on a nanosecond scale.

I can't wait to get older and see an all-atom bacteria simulation!