Questions: How is the brain wired, how did the universe start, how does matter interact at the quantum level, how does the human visual system work, how can we prevent heart attacks?  Answer: Hanspeter Pfister. Yes, that's the answer. Hanspeter wrestles with some of the largest questions faced by scientists and researchers from his post as a computer scientist at Harvard University. And, as you can imagine, tackling those questions require some heavy compute power.  

For instance, a project known as Connectome (so dubbed because this project is on par with the Humane Genome Project) is trying to map the wiring diagram of the brain down to individual neurons and synapse connections. Hanspeter estimated that mapping just 1 cubic millimeter of brain tissue would require 1.5 Petabytes of storage. "This will be an exascale-size data project," he said, echoing some of yesterday's comments envisioning the roadmap to make this a reality.    

He also discussed a radio astronomy project known as the Murchison Widefield Array attempting to understand what happened in the time frame 300,000 years after the Big Bang to 1 billion years after the Big Bang. Apparently, and I was unaware of this, not much is known about that stretch of time. In conjunction with the Harvard Center for Astrophysics and others, they are building in middle-of-the-middle-of-nowhere Boolardy, Australia, an antenna array and "supercomputer" center (in a temporary trailer) that can only pull 20kW to do all of the necessary computations. With such a power constraint, Hanspeter opted for a GPU cluster to churn the data.    

Hanspeter sees a growing necessity in moving data crunching capabilities closer to the source of the data collection to eliminate the bandwidth restrictions of transmitting data (whether that be via data pipes or via FedEx). Other benefits Hanspeter gets from bringing an inexpensive computational source such as GPU clusters closer to the data: Creation of a feedback loop and an ability to scale systems.  

The projects he discussed on Thursday were of immense scope, scale and complexity. Utilizing new computing techniques to tackle them has provided him with new opportunities, but he did highlight some challenges with using GPUs for high throughput computing, namely the need for higher-level programming models, scalable programming models, and plug-and-play parallel high-throughput I/O. Scientists, for example, want to use the programming languages they already know so he said more accommodation of domain-specific languages would be helpful.