|Jen-Hsun Huang shows off “Fermi” at today’s GPU Technology Conference. Source: nVidia|
SAN JOSE, Calif. — nVidia kicked off its GPU Technology Conference today with the formal introduction of its next generation in graphics architecture, codenamed “Fermi,” that CEO Jen-Hsun Huang called “the soul of a supercomputer in the body of a GPU.”
The first board, shown by Huang to the audience, is still fresh off the line and there is no release date yet. In fact, Huang told the gamers in the audience, nVidia’s (NASDAQ: NVDA) primary audience for most of its career, that they would have to wait for details on new graphics boards.
Indeed, the Fermi architecture, which has been the subject of rumors of delays, is no longer just concerned with gamers.
“Instead of a GPU that has been extended to do supercomputing, we took the risk to come out with a new architecture that is designed to be a computer first and treat graphics and parallel computing as equal citizens,” Huang told the crowd here at the show.
Fermi is a beast of a chip, with three billion transistors. Intel’s largest chips, the quad-core Itanium “Tukwila” and eight-core Nehalem-EX, weigh in at just 2.3 billion transistors, and that’s enormous by Intel standards.
nVidia’s new offering has twice the number of cores — 512 total — and eight times the double-precision, floating-point performance as the previous generation, something important in areas like particle simulations and other physics simulations the Fermi is designed to run.
The GPU has six 64-bit memory partitions for a 384-bit memory interface supporting up to a total of 6 GB of GDDR5 DRAM memory. It also comes with native C++ support through unified address space and ECC memory support for catching soft errors.
Jon Peddie, president of Jon Peddie Research, said nVidia is growing up from its video game roots out of necessity.
“nVidia has been a company that has exhibited phenomenal growth for the past six years and in the process has just about saturated all the markets they’ve entered, so they have to find new growth areas,” he told InternetNews.com after the keynote speech. High-performance computing “is a wide-open market area for it and they’ve shown they have the engine to participate.”
Huang said at a press conference after that launch that Fermi would likely ship in the first quarter of 2010, but not in high volume.
“Sexy” developer technology
Following up a demo of Fermi against the last-generation Tesla, Huang then showed Nexus, a source-level debugger that works inside of Microsoft’s Visual Studio developer tools to debug heterogeneous computing applications. This means apps written to support both a CPU and GPU can be debugged in the same place.
“This is a real breakthrough. This is a technology only a real developer would find sexy,” Huang joked. The beta version comes out Oct. 15.
Jeffrey Nichols, associate laboratory director for Computing and Computational Sciences, Oak Ridge National Laboratory, joined Huang to discuss the Labs simulation work with Jaguar, the second-fastest supercomputer on Earth as of this past summer.
Jaguar is built on AMD Opteron processors, but Nichols talked about how he plans to leverage the increased power of Fermi for developing climate models that break down regional weather patterns. Until now, supercomputers have not used GPUs, so this would reflect a change in computing strategy.
Fermi is also a change in strategy for Vidia, too. GPUs are going to do more than the texture mapping they have been known for, Huang promised.
“This year will bring real-time ray tracing with GPUs. Ultimately, we’re trying to capture and interact with a dynamic world. The more interactive and dynamic it becomes, the more complex it becomes,” he said.
What he is aiming for is to make virtual worlds more complicated and ultimately more unpredictable, so no two interactions are the same. In a fluid dynamics demo, he showed how water motion was different each time the simulation ran due to random interactions of the waves, something that was impossible in previous generations.
Another non-gaming speaker was David Robinson, CEO of Techniscan, the maker of ultrasound equipment using Tesla GPUs to process mammograms. He joined Huang on stage to discuss some unpleasant facts about breast cancer screenings: 30 percent of scans miss tumors, some women can’t be properly scanned by CT machines, and processing the image can take an hour.
Using a Tesla machine, TechniScan is already able to process nine million Voxels
“The idea is that we can detect more cancers, smaller cancers, cancers in women that mammography doesn’t serve well, that’s exactly our goal,” he said.