Nvidia's efforts to push graphics chip technology into mainstream computing received a significant boost May 18 when IBM announced that the newest version of its iDataPlex server will run on both traditional computing chips and GPUs from Nvidia.
The announcement marks the first time a top-tier systems vendor has made the move to offer a hybrid CPU-GPU server in the HPC (high-performance computing) space.
Nvidia officials said having IBM adopt Tesla 20-series GPUs in its iDataPlex dx360 M3 system is a validation of both Nvidia and of the vision the company is promoting.
"Having IBM [put GPUs into the server] means that anyone considering a high-performance computing system will now have options," Sumit Gupta, senior product manager for Nvidia's Tesla business, said in an interview, noting that IBM and Hewlett-Packard dominate the Top500 list of the world's most powerful supercomputers. "This really makes GPUs a mainstream [technology]."
Nvidia has led the effort to bring GPUs into mainstream, general-purpose computing. Company officials unveiled the Tesla 20-series in November 2009 at the Supercomputing show in Portland, Ore. Based on Nvidia's new "Fermi" architecture, the Tesla 20-series graphics chips are meant to be used for general-purpose computing, offering the performance of traditional CPU-based clusters, but at a 10th of the cost and a 20th of the power.
Nvidia officials began promoting the shift in 2006, when the company rolled out the CUDA architecture. Several smaller HPC server makers, including SGI, Appro and Super Micro Computer, are already using Nvidia GPUs in some of their systems. IBM will be the largest OEM to do so.
In a video presentation, Scott Denham, a member of IBM's Deep Computing technical team, said the IT industry needs to find a way to meet business demands for more performance from their systems, but at less cost.
"Many of our customers are constrained by the amount of power and cooling they can get into their existing facilities," Denham said. "They have a demand for a much higher level of computation to implement advanced algorithms."
The only real options for doing that are to build larger buildings, bring in larger power supplies or find another way of doing things.
"GPU technology is an ideal way to get this much higher level of computational ability in this same power footprint as the existing systems while increasing throughput four to 10 times," Denham said.
The iDataPlex dx360 M3 server is powered by two Intel CPUs, and two GPU cards are attached. The architectures can "add a tremendous amount of compute power to the existing infrastructure," he said.
Nvidia officials are saying in order to keep computing capabilities improving at the same pace they have for the past few decades, there has to be shift away from serial CPUs and toward parallel processing found in GPUs.
Intel and AMD are looking to keep making significant gains in performance through small chip sizes and more processing cores. However, in a column on Forbes.com in April, Bill Dally, vice president and chief technologist at Nvidia, said the power needs of compute chips from Intel and Advanced Micro Devices are making it difficult to sustain Moore's Law.
"It is past time for the computing industry-and everyone who relies on it for continued improvements in productivity, economic growth and social progress-to take the leap into parallel processing," Dally wrote.
IBM's Denham agreed.
"The future of computing in high-performance computing space [is] moving toward this hybrid model where we have a combination of standard [CPU] cores and very high-performance [GPU] cores, hundreds per chip," he said, adding that parallel computing is the answer. "Parallelism inside the chip is the most effective way to do that."
AMD, on the strength of its ATI unit, is looking to closely integrate CPU and GPU technologies on a single chip, part of its Fusion initiative. Intel also is enhancing the graphics capabilities of its CPUs.