Jen-Hsun Huang thinks GPU computing is just starting to hit its stride. Kicking off the annual GPU Technology Conference in San Jose, Calif. on Tuesday, Nvidia's president and chief executive promised that the company's new Kepler architecture and thriving CUDA ecosystem would continue to "democratize high-performance computing as we know it."
Nvidia's evolution from a company that largely catered to hardcore games to one that's got a finger in everything from scientific computing to smartphones may be one of the most under-reported stories in technology. Back in 2008, when the graphics chip maker first introduced its proprietary CUDA programming language, there was a single Top 500 supercomputer that used its graphics processors.
Today there are 35 supercomputers built around Nvidia's Tesla GPUs, Huang said. The parallel processing capabilities of CUDA-powered GPUs are being used for everything from dendrite simulations and air traffic modeling to adaptive radiotherapy and collective behavior research, he noted.
Kepler, Nvidia's latest GPU architecture and the successor to Fermi, is "a GPU that will fundamentally reinvent computer graphics yet again," he said.
"It is the best GPU we've ever built. It is the most energy-efficient GPU we've ever built, and as you will learn today, it is also the most feature-packed GPU we have ever built," Huang continued.
Nvidia's first Kepler products included GeForce-branded GPUs for desktops and laptops such as the top-of-the-line GeForce GTX 690 card, which sported a pair of Kepler GPUs.
"Some people say it's the most beautiful graphics card they've ever seen, some say it's the best weapon they've ever had. I say, why bicker. We're lovers not fighters," Huang joked.
To highlight just how powerful the new architecture is, Huang served up impressive demos of Kepler GPUs shattering crystals and successfully performing a combination of ray-tracing light and fluid simulation—all in real time.
The Kepler architecture isn't reserved for GeForce. It will eventually waterfall down to Nvidia's Tegra product line for consumer mobile devices and up to the Quadro family of professional workstation products, but before that happens, Nvidia is bringing Kepler to Tesla, a powerful line of GPU products for the HPC and supercomputer markets.
On Tuesday, Nvidia announced the availability of the Tesla K10, which offers three times the single-precision floating point performance of Fermi-based Tesla cards, plus 1.8 times the memory bandwidth. Nvidia plans to follow up in the fourth quarter with the Tesla K20, which will treble the double-precision performance of Fermi and introduce the company's "Hyper-Q" technology, which for the first time adds multiple work queues from a CPU, resulting in a much higher GPU utilization rate than was possible with single-queue architectures like Fermi and its predecessors.
Huang also gave a glimpse of a future where GPUs will finally join CPUs as virtualized tools for cloud computing, thanks to the integration of virtualization technology in Kepler in combination with ultra-low latency remote display technology, and the ability to better scale out Kepler GPUs in data centers thanks to improvements to the chip's energy efficiency.
"We're taking the GPU into the cloud. Kepler is the first GPU designed for the cloud," he said.
For more from Damon, follow him on Twitter @dpoeter.
