The NVIDIA GPU Technology Conference (GTC 2010) kicked off today with their energetic CEO Jen-Hsen Huang presenting the keynote where he spoke about four things- NVIDIA’s focus, CUDA’s progress and Status, new announcements and finally a peak into the future.
Jen-Hsen stated that NVIDIA’s focus continues to remain on three key areas- visual computing, parallel computing and mobile computing with their Quadro, Tesla, GeForce and Tegra brands. He stated that mobile computing is an area where great contributions to computing will happen and described it as the first computer that is equipped with all kinds of sensors such as camera, microphone, GPS and accelerometers making it context and situation aware. The Tegra brand he described, is SOC (System on a Chip) and based on the ARM architecture which is the fastest growing CPU in the history and is the ISA of choice for mobile computing. He said that NVIDIA is super excited about the combination that a Tegra device will offer along with Tesla based cloud computing- calling it a Magical experience (Is that the new buzz word now?)
He continued evangelizing tessellation which is the key marketing point in NVIDIA’s Fermi based GPUs such as the entire GTX400 family. Tessellation basically creates geometric fidelity from complicated Shaders, displacement map and a patch. He stated that Fermi is six times faster at geometric processing than their previous GPU- a leap that has never been realized in any consecutive generation of GPUs by NVIDIA.
He then proceeded to show the demo of HAWX 2, the upcoming sequel to Ubisoft’s game. Each level of this game is about 128 square kilometers and each polygon comes all the way down to two meters so within each level, four billion triangles are represented. It looked pretty amazing, especially since it was showcased in 3D.
Next he spoke about CUDA parallel computing architecture that was introduced in NVIDIA GPUs about four years back. He said that NVIDIA realized that the GPU could be used outside the graphics territory for General purpose computing and as a result CUDA benefited from the billion dollar budget that is allocated towards building GPUs. Incorporating CUDA in every single GPU was a strategic and expensive decision for NVIDIA but he thinks the benefits pay off. He mentioned the different applications that CUDA is being used for by developers such as research on biology, finance, product design and astro-physics.
He said NVIDIA realized that High-Performance Computing is becoming the third pillar of science and GPU computing makes super computing much more affordable and accessible. Through a graph, he showed that in just ten years from now, we would lose a hundred times the performance benefit if we continue to use CPUs in areas that GPU could replace. He said CUDA usage and adoption is coming along quite nicely between last year and this year with the number of SDK downloads doubled.
Jen-Hsen had quite a few announcements to make for CUDA. First of all all, NVIDIA announced CUDA-x86 in partnership with PGI that allows CUDA applications to be complied and deploy on CPU clusters and not just GPU clusters. He then announced CUDA support for a number of applications such as MatLab, Amber 11, ANSYS and 3DS Max – I haven’t used any of these applications but these apps are pretty much the standard in their respective industries. And from the announcements, it seems that CUDA can speed up parts of these applications from 2x to 50x. That is a pretty big deal for anyone who uses these applications- probably a bit like when you first saw Quake running on accelerated hardware.
Following these announcements, Jen-Hsen invited a team of two from Adobe on stage to showcase the Plenoptic lens- a technology that puts a whole bunch of lenses on one sensor. So instead of one ray per pixel, you take multiple smaller rays per pixel using a micro-lens array- basically capturing multiple but smaller sized images instead of one big image. And then using Software, you can pretty much decide what you want to focus in your picture. The following video will explain the technology and what you can do with it. Arguably, this was the coolest demo at the show yesterday:
Finally, Jen-Hsen closed the keynote by doing something NVIDIA has never done before- show a roadmap of their upcoming GPUs- the Kepler based on 28 nanometer and expected in 2011 and following that, Maxwell in 2013.
Jen-Hsen showed a performance comparison using performance per watt and comparing it to Fermi, Kepler should offer about 2-3 times perf/watt. Maxwell on the other hand, expected in 2013 will offer almost 8 times the perf/watt than Fermi.