If there’s one take away from this week’s NVIDIA GPU Technology Conference (GTC), it’s that GPU computing has grown up. Having been to last year’s event, it’s amazing to see how many more academic researchers and companies are taking the technology seriously in 2010. The exhibition hall was twice the size of GTC in 2009, enough to accommodate the 100 or so vendors plying their GPGPU wares. As NVIDIA CEO Jen-Hsun Huang said in Thursday morning’s fireside chat session, “This is the year when applications developed on GPU computing go into production.”

There was so much activity centered on technical computing at this year’s event that at times it seemed like a CPU-less version of November’s Supercomputing Conference. That was also reflected in the exhibitor list, which included HPC stalwarts like IBM, HP, SGI, Dell, Appro, Supermicro, Microsoft, The Portland Group, Platform Computing, Mellanox, T-Platforms and at least a dozen others.

Application areas like seismic exploration, weather modeling, computer vision, and medical imaging are latching onto this technology quickly. Just slightly further behind are domains like biomolecular modeling, which appears to be ripe for the GPU. The Wednesday keynote by Dr. Klaus Schulten, a computational chemist at University of Illinois, Urbana-Champaign, highlighted some early benefits in this area.

Schulten and his team at UI have started applying GPU acceleration to a range of molecular simulations. In his work, Schulten is employing GPGPU technology to develop the concept of a “computational microscope,” which is designed for nanoscale examination of biomolecules and cells. This virtual microscope consists of basic chemistry and physics algorithms, NAMD software (which will soon offer a GPU port), and supercomputing hardware.

One application that Schulten talked about was modeling the flu drug Tamiflu to determine how the H1N1 (“swine flu”) virus developed resistance to it. He’s also using the technology to study such phenomenon as virus infections, how proteins are synthesized, the mechanism of photosynthesis, epigenetics, and quantum chemistry. Some of the work is being accomplished on GPU workstations, but the larger models use NCSA’s Lincoln supercomputer, a heterogeneous cluster constructed from Dell PowerEdge servers and S1070 Tesla servers. Speedups on applications varied, the best being the quantum chemistry application. In that case, a simulation run that took a day with a CPU, took just a minute on the GPU platform.

There were a couple of sessions on the military applications of GPU computing, which looks to be a lucrative area for this technology. One presentation, hosted by EM Photonics, illustrated how GPGPU technology is being employed to accelerate compute-intensive applications in this domain. For example, an advanced image processing application was able to enhance long-distance photographs blurred by atmospheric distortion. GPU acceleration made it possible to perform this digital enhancement in real-time, opening up new applications for warfare and security operations. Other apps include electromagnetics simulations and CFD — the latter being used to simulate aircraft landings on carriers. Depending on the military scenario, the GPU platform could be a desktop machine, an embedded system, or a cluster.

Other GPU computing applications that got some exposure at GTC this year are business intelligence, complex event processing, and speech recognition — three areas that up until now would not have been associated with graphics processors. And of course there were a plethora of esoteric research applications, for example, Using GPUs for Real-Time Brain-Computer Interfaces — something that would have come in handy at GTC this week, give the overload of sessions, posters, exhibits, and after-hours partying.

This also looks to be a breakout year for ISV support of GPGPU in HPC. At the event, ANSYS announced it would be incorporating GPU acceleration into its popular engineering modeling and analysis solution, ANSYS Mechanical. That product is slated for release later in the year. And although SIMULIA and Livermore Software Technology Corp. (LSTC) made no formal announcements this week, two GTC presentations on Thursday suggest they also will be bringing out GPGPU-support for their flagship products (Abaqus FEA and LS-Dyna, respectively) within the next few months.

Even though GTC was more about developers and applications, there were a few sessions highlighting some of the larger GPU supercomputers deployed, or about to be deployed. In this latter category is TSUBAME 2.0, Tokyo Tech’s next-generation 2.4 petaflop super, which will be stuffed to the gills with 4,244 Tesla M2050 GPUs. In Tuesday’s presentation by Satoshi Matsuoka, he spotlighted some of the cutting-edge apps that will be running on the new machine. This includes ASUCA, Japan’s next-generation weather forecasting code that has been completely ported to the GPU (and reportedly took a year to do so). The result is that they will have a weather modeling application that is faster than real-time and works at resolutions of 0.5 km. According to Matsuoka, TSUBAME 2.0 is installed and undergoing stress tests, and will be formally announced in early October — so expect more coverage to follow.

If 2.4 petaflop supers don’t impress you, you’ll just have to wait a bit. Thanks to a brief peek at NVIDIA’s roadmap on Tuesday, the next generation of NVIDIA GPUs, Kepler, is slated to arrive in 2011. As Jen-Hsun Huang noted, “GPU computing is just starting. It’s nothing compared to what you’re going to have in a couple of years.”