Two more legendary scientists, Johannes Kepler and James Clerk Maxwell, join the others whose names have been inducted into Nvidia’s computing architecture repertoire.
On Tuesday, September 21, Nvidia co-founder, president, and CEO Jen-Hsun Huang, took the
stage to address the crowd at the annual GPU Technology Conference. The backdrop is HAWX 2,
Ubisoft’s upcoming game, used as demonstration of GPU-powered tessellation.
The company’s current-generation GPU architecture, Fermi, was named after the Italian-American physicist Enrico Fermi. It represented a major technological leap—“six times faster than the previous generation,” recalls Jen-Hsun Huang, Nvidia’s cofounder, president, and CEO. At the recent GPU Technology Conference (GTC) in San Jose, California, Huang introduced Fermi’s successors: Kepler and Maxwell.
Fermi-class Nvidia cards could process up to 1.5 double-precision giga-FLOPS (DP GFLOPS) per watt. Under Fermi’s immediate successor, Kepler, Nvidia expects the cards to reach somewhere between four to six DP GFLOP per watt. By the time Maxwell comes along, the numbers could reach 14 to 16 GFLOPS per watt, according to Huang’s ambitious road map.
Ken Pimentel (Autodesk Media and Entertainment), Michael Kaplan (Mental Images), and Jen-Hsun
Huang (Nvidia) get ready to demonstrate the use of the Iray renderer in 3ds Max.
Tessellated Realities, Massive to Mobile
The big idea for this year, according to Huang, is “tessellation.” He explains, “Tessellation simply takes a patch, a displacement map, a shader, and creates amazing geometric fidelity where there wasn’t [any] before.”
Fermi’s geometry processing horsepower plays a critical role in enabling tessellation, Hung points out. The use of tessellation to handle massive data sets will become evident in Ubisoft’s upcoming game, HAWX 2, an aerial squadron warfare simulator. “Each level of the game is approximately 126 square kilometers,” explains Huang. “Each polygon comes all the way down to two meters. So if we render each level, it will represent four billion triangles.”
Another demonstration, a 3D fly-through environment titled Endless City, was procedurally generated with more than 1000 light sources. Running at about 1.3 billion polygons per second, the visualization session was made possible only with the parallel computation architecture CUDA brought to the scene, Hung maintains.
Hung called high-performance computing (HPC) “the third pillar of science.” Once considered an IT resource exclusive to large enterprises, HPC—specifically GPU-powered HPC systems—are about to make applications more affordable. “A tipping point has been reached,” says Huang. “Scientists and engineers around the world want to have access to HPC.”
Endless City, a visualization sequence showing how tessellation can be used to drive a procedurally
generated cityscape (compare its look without tessellation on the left and tessellated look on the right)
If HPC is for large-scale computation jobs, such as realistic simulation and real-time visualization, mobile computing may be the opposite end. But the challenges it presents to content providers are no smaller just because the display screen is.
“Mobile computing isn’t just about a computer you can take with you … this is the first computer that’s equipped with all kinds of sensors: camera, microphone, GPU, accelerometer,” notes Huang. “This is the first computer that’s context-aware, situation-aware. Who knows? Maybe one of these days it’ll be self-aware. Because of all these sensors, mobile computing will be able to achieve and deliver the kind of magical experience that wasn’t possible before.”
With four distinct brands, Nvidia hopes to capture the entire spectrum of applications: GeForce for gaming and consumer market; Quadro for professional market; Tesla for HPC market; Tegra for mobile market.
From Analysis and Rendering to Calculation
There are two things Nvidia did right with CUDA, Fermi’s programmable environment, Haung reflects. “We recognized that we don’t want to replace the CPU; we just want to add to it … [We rode on] the back of the high-volume GPU business we had, GeForce,” he outlines.
Endless City, a visualization sequence showing how tessellation can be used to drive a procedurally
generated cityscape (compare its look without tessellation on the left and tessellated look on the right)
The growing number of applications coming online is evident in Nvidia’s download statistics. In 2009, the CUDA developer kit was downloaded 293,000 times. This number jumped to 668,000 in 2010. Recent software developers joining the CUDA base camp include Matlab, often described as Excel for engineers; Ansys, a widely adopted simulation software maker; and Autodesk, for its 3ds Max digital content creation package.
In Autodesk 3ds Max, users will soon have the option to render their scenes on the GPU using Mental Images’ Iray rendering engine. Soon they will also be able to access and visualize uploaded 3D scenes remotely from a browser.
Parallel Bites
One way Nvidia plans to promote CUDA further is with its partnership with The Portland Group (PGI). At GTC, Nvidia announced that PGI is developing “a CUDA C compiler targeting systems based on the industry-standard, general-purpose 64- and 32-bit x86 architectures.” PGI plans to demonstrate the compiler at the Supercomputing Conference 2010 (New Orleans, Louisiana, November 13-15).
“The hundreds and hundreds of man-years that go into something that’s really difficult to describe—that’s what we call [computing] architecture efficiency,” says Huang as he got ready to name Kepler and Maxwell.
The smoke volumes here are generated through computation and fluid dynamics simulation,
powered by the GPU.
If you insist on evidence, you may launch the System Performance Tab in the Windows task manager to view your GPU usage. But that’s not exactly the best way to understand the speed and performance made possible by the GPU. You’ll have to experience it in the animated narratives, physically accurate crash tests, and digital photos virtually indistinguishable from reality.
Kenneth Wong is a freelance writer who has covered the digital video, computer gaming, and CAD industries. He can be reached at
Kennethwongsf@earthlink.net
.