| by
GPU Conference PresentationsI mentioned previously about a GPU conference that discussed the implications of the technology for doing scientific research. There are certain classes of problems that GPU's are especially suited for and they offer a speed up when compared to CPU's. As an example, researchers have recently developed a relatively inexpensive 13 gpu "supercomputer" with about 12 teraflops of computing power for scientific problems. GPU's have been rivaling the complexity of intel's most advanced technology and 3 billion transistor gpu chips will probably hit the market shortly. Nvidia also believes that they can reach 10 billion transistors easily. With this speed up of processing power coupled with machine learning, we will be able to learn more about the brain than ever before. While it's probably somewhat facile to make a blanket statement that computing power is increasing exponentially, there are still some interesting exponential trends in the field that will likely continue for at least the next 5 or ten years.

Nvidia has put media from that 2009 conference online and several of them are related to neuroscience. The company Evolved Machines is "pioneering the reverse engineering of brain circuitry to build intelligent machines". An audio talk can be found here (6.1 MB).

"Reconstructing the Brain: Extracting Neural Circuitry with CUDA and MPI" is a 37.6 MB video presentation (download here). The following is an excerpt about that video;
In this talk we will present our insights and lessons learned in using CUDA to reconstruct neural connections in high-resolution EM data. We will present technical details and non-trivial issues regarding the implementation of NeuroTrace, our system for semi-automatic segmentation and interactive visualization of terabytes of EM image data. The segmentation method is based on a sequence of 2D level set segmentations of cell membranes integrated with an image correspondence energy for robust transition between consecutive slices and a weighted path extrapolation method to trace a 3D centerline of a neural pathway along non-axis aligned slices.
Optimizing Ion Channel Kinetics Using A Massively Parallel Genetic Algorithm on the GPU (26.4 MB video presentation);
Voltage-gated ion channels effect the integration of information in many neurons. Some neurons express over 10 voltage-gated channels that turn information processing into a highly non-linear affair.
The currently popular analysis techniques suffer from various shortcomings that limit the ability of the researcher to rapidly produce physiologically relevant models of voltage-gated ion channels.

To solve this computational bottleneck we have been converting our optimization algorithm to work on a GPU using CUDA. We have succeeded to parallelize the process on a GTX 295 giving a speed increase of roughly X100 over that of the CPU.
Medical Image Registration with CUDA (37.6 MB video presentation);
Speedups of up to 750 times were obtained as compared to code in daily use at Addenbrookes Hospital and Bio-Medical Campus. Some very recent results are shown in the figures. This work is of direct application in both research and clinical practice. A particular application is voxel based MRI morphometry in humans and in animal brains.
High-Throughput Science (keynote speech);
How did the universe start? How is the brain wired? How does matter interact at the quantum level? These are some of the great scientific challenges of our times, and answering them requires bigger scientific instruments, increasingly precise imaging equipment and ever-more complex computer simulations.
The rest of the presentations can be found here. They cover a wide variety of topics.


Add a Comment GPU Conference Presentations