Nvidia’s GPU Technology Conference, Day 2

Day Two – focusing on research and the emerging companies

It was warmer in San Jose the next two days and no jacket was required any longer as the conference also got into full swing.  We noted that the first day’s events was followed by a nighttime pub crawl through downtown San Jose to raise money for charity.  Quite a few people were quite tired – but happy – even the morning after and no enthusiasm appears to have been lost for the GTC whatsoever.

Day Two Keynote: Research

Harvard Professor Hanspeter Pfister began a fascinating 90 minute talk at 9 AM on how GPUs are being used to answer some of science’s thorniest problems – like how the brain is wired, how the human visual system works, as well as how the universe began.  He also discussed the very pragmatic subject of preventing heart attacks with real-time advanced MRI-type scanners that will likely become hand-held miniature diagnostic devices that should evolve into the “tri-corder” of the future.  What was unusual is that he was able to make the subject of ‘matter interacting at the quantum level’, interesting to experts as well as to non-technical viewers in the audience.  That is a rare gift.

The Connectome Pipeline

Let’s cover some of the subjects that he presented that demonstrate the need for GPU parallel computing and its incredible advantages over CPU supercomputers. One project has to do with finding out how the brain is wired and it is massive, rivaling the Human Genome project in every way.  It is called finding the Connectome pipeline and this diagram is the simplest representation of what they want to achieve as they map the synaptic connections.

They are using rat brains in their research and the idea is to slice the brain tissue into extremely thin but long slices – much as one uses an apple peeler or a lathe.

Next it is imaged with electron microscopes.  Also, notice the 40 GB of data that is required to image such tiny sections and the fact that they need many samples to make a 3D image of their overall sample that will put a massive strain on the fastest supercomputer.  We are talking about 1 cubic MM requiring 1.5 PB of data to manipulate!

They certainly aren’t done yet.  After they have properly imaged the section of rat brain sample they took, they next need to reconstruct it and map it.  Here we see reconstruction and active brain ribbons.

Again, this takes massive computational power as they do an elliptical approximation of ribbons with a compression ratio of 1000 to 1.  Here we see the active ribbons processing on the GPU contrasted with the CPU supercomputer.  A twenty-three times speedup is far beyond extra-ordinary.  It completely changes the scientist’s work for the better.

They already have some incredible results with these multi-scale images.  What is making it even more interesting is that the test rats are genetically bred so that their tissue now show up in colors under ultraviolet light so that their brains can be visualized and mapped easier.  They call it the “Brainbow”.

One of the slides half-humorously compared the brain’s wiring to that of a PC:

The future is wide open to this kind of GPU parallel computing that is so much faster than what they had before using only the CPU.