NVIDIA’s GPU Technology Conference, held last week in the lovely city of San Jose, is remarkably different from the company’s other events. This conference is not aimed at promoting the next big thing in consumer electronics, or a celebration of gaming – as was October’s GeForce LAN –, but rather a preview of the possible applications of NVIDIA’s new GPU in applied science and computing.
To the consumer, the GPU arms race is about which company will make games look better at a higher resolution. Many know NVIDIA for just that: the company that enables gaming. However, the rapid advancement in GPUs that brought us to Kepler from Fermi is also, in the words of NVIDIA CEO Jen-Hsun Huang, the “democratization” of High Performance Computing (HPC).
The chip that NVIDIA believes will be an agent of the “democratization” of HPC is the Kepler based GK110 GPU found in the upcoming Tesla K20 card. As discussed in previous coverage on Hardware Canucks, the competitive advantage NVIDIA has packed into these chips is Dynamic Parallelism and Hyper-Q.
While in the world of NVIDIA’s Fermi chip-of-yore, only the CPU was able to introduce kernels; the GK110 introduced the ability for kernels into introduce child sub-kernels for the GPU to work through while the CPU is still occupied.
When combined with the move to the 28nm fabrication (as opposed to 40nm for Fermi) process for Kepler, NVIDIA is able to dramatically increase efficiency by packing in more cores – 192 vs. Fermi’s 32 – into the chip while reducing the clock speed of each individual core. Because of the enhanced parallelism capabilities of the chip, NVIDIA is able to boost performance while lowering power draw and overall heat output.
The hurdle in HPC that NVIDIA is attempting to cross with the release of the Kepler powered GK110 is the CPU/GPU bottleneck. To jump this hurdle, NVIDIA is introducing Hyper-Q, which allows for message passing interface (MPI) commands to be sent between CPUs and GPUs in CPU/GPU clusters in parallel thus reducing the time it takes for the computer to run through the code on hand.
What is the end result of the GK110? 1TFlop of double-precision floating point processing power (or so NVIDIA claims). The question is, what are the ‘killer applications’ of this kind of power?
Emerging Companies Summit: GPUs Growing Companies
The Emerging Companies Summit was a conference within a conference at GTC. While GTC was focused on the GPU as an emerging trend in computing, the Emerging Companies Summit was focused on companies using the GPU as a vehicle of growth.
The star of the summit was Gaikai, the company NVIDIA is teaming up with to power the GeForce GRID. While Gaikai’s model isn’t inherently new – as OnLive currently offers such a service – the use of Kepler GPUs in the datacenters that power Gaikai means they, or so they claim, can offer games at a latency which is similar to set top consoles or P.Cs.
Gaikai’s presentation was scant on technical details; their CEO only rehashed what was already said by NVIDIA’s Jen-Hsun Huang. What really wowed the audience was that the technology is already partially available: while no GRID data centers have been deployed, LG’s Smart TVs are already shipping with Gaikai’s app.
This is a company to watch in the next 2-3 years. If NVIDIA and Gaikai do it right, they could very well be the ‘iTunes’ of gaming.
Moving Towards a Holodeck: RealView Imaging
Israeli firm RealView Imaging hasn’t yet built a holodeck, but what they displayed at the Emerging Companies Summit could be seen as a step in that direction.
RealView Imaging’s technology projects, in 3D, an image that can be manipulated by the user simply moving their hands. The company is marketing this towards the medical market, and when combined with next generation MRI scanners the possibilities are enormous.
Eventually RealView Imaging’s technology will make its way to industrial and architectural production and design.
The ‘killer app’ for this technology can be found in an application of the research presented in the second keynote of the conference on the Principles of Collective Behavior by Ian Couzin, a researcher from Princeton.
Professor Couzin’s keynote described his research – powered by Kepler GPUs – on collective behavior in hordes, be it a school of fish, a flock of birds, or people in urban environments. Combine this with Realview Imaging’s 3D projections, and you have an application that will allow an architect and engineer to simulate how people navigate through the labyrinth of a potential new urban development, or allow those charged with disaster management to simulate the evacuation of a building or subway system.
This technology is inspiring, and the potential applications for a better designed world are endless.
Conclusion: Better Living Through CUDA
As emphasized repeatedly during the conference keynotes, with Kepler High Powered Computing just got amazingly affordable. The computing power that was once the domain of major universities and big corporation’s research arms a decade ago is now in the tens of thousands of dollars price point. It’s accessible for the masses.
CUDA, the SDK that NVIDIA introduced in 2006, first allowed developers to exploit the power-efficent massively parallel computing that is at is now at its zenith with Kepler. Because of this advancement, NVIDIA now projects that nearly 75 percent of HPC customers will use GPUs for general purpose computing in 2014. The fastest computers now come with the likes of an NVIDIA Tesla, not some purpose built proprietary hardware.
Next month, a suburb of Seattle, AMD will host the AMD Fusion Developers Summit –its attempt at a GTC. They’ve seen NVIDIA’s offerings. The question is, can they compete?
As NVIDIA’s CEO Jen-Hsun Huang emphasized during his ‘fireside chat’ that kicked off the Emerging Companies Summit, NVIDIA simply wants to be a “company that matters”. A GPU arms race simply to create better looking games is trivial pursuit, though using those same refined architectures to solve the world’s problems through hyper efficient computing.
Tags: GTC 2012