Nvidia and Advanced Micro Devices are pushing the capabilities of their respective graphics cards as they look to capture larger shares of the growing interest in GPU accelerators in the high-performance computing space.
At the SC ’13 supercomputing show in Denver Nov. 18, Nvidia officials unveiled the Tesla K40 GPU accelerator, which the company said will offer twice the memory—at 12GB—and 40 percent the performance of its current K20X accelerators. The chip, aimed at such workloads as big data and scientific calculations, also offers 10 times the performance of systems using only CPUs, according to Nvidia officials.
Nvidia’s announcement came four days after AMD rolled out its FirePro S10000 12GB Edition graphics card, which is aimed at high-performance computing (HPC) and big data tasks and marks the vendor’s most significant push into the high-end computing field with its GPU technology.
“Our compute application customers asked for a solution that offers increased memory to support larger data sets as they create new products and services,” David Cummings, senior director and general manager of professional graphics at AMD, said in a statement. “In response, we’re announcing the AMD FirePro S10000 12GB Edition graphics card to meet that additional memory demand with support for OpenCL and high-end compute and graphics technologies.”
Both Nvidia and AMD for the past several years have been promoting GPUs as accelerators for use in high-end systems running HPC workloads. The GPUs enable OEMs to build systems that can improve performance and handle more parallel-processing workloads while also driving up the power efficiency, an increasingly critical metric in highly dense data centers.
Interest in the HPC space in GPU accelerators and coprocessors—like Intel’s x86-based Xeon Phi many-core chips—continues to gain momentum. According to the organizers of the Top500 list of the world’s fastest supercomputers, 53 systems on the list are using either GPU accelerators or coprocessors—38 of which are using Nvidia GPUs, and another two using AMD’s. Thirteen are using Xeon Phi coprocessors.
More than 240 applications can leverage GPU acceleration, according to Nvidia.
Sumit Gupta, general manager of Tesla accelerated computing products at Nvidia, said interest in GPU accelerators has been strong over the past few years.
“No one expected that kind of adoption,” Gupta told eWEEK, adding that the kinds of use cases for systems with GPU accelerators range greatly, from use on Dutch Navy warships to systems that help develop new shampoos. End users “keep coming up with cool use cases.”
Nvidia’s new K40 GPU is being embraced by a range of OEMs—such as Dell, IBM, Hewlett-Packard, Cray, SGI and Bull—that will being rolling out systems leveraging the graphics cards.
“NVIDIA’s accelerators enable our customers to realize significant improvements in processing performance,” Bill Mannel, general manager of compute at SGI, said in a statement.
Nvidia, AMD Bulk Up Their GPU Accelerators
The new GPU, based on the Kepler architecture, includes Nvidia’s GPU Boost technology, which can help businesses get more performance by leveraging power headroom, according to the company. GPU Boost can help systems become 25 to 40 percent faster, Gupta said. The chip exceeds other accelerators for both single-precision (4.29 teraflops) and double-precision (1.43 teraflops) peak floating-point performance, Nvidia officials said.
Along with the 12GB of GDDR5 memory—which enables organization to process datasets that are twice as large as those handled by 6GB memory—the K40 also includes 2,880 CUDA parallel processing cores that offer 10 times the acceleration of systems that use a CPU alone, and PCI-Express Gen-3 support, which offers twice the data movement acceleration as the PCIe Gen-2 technology.
The K40 GPUs are available now, and users can try out the accelerators for free on remotely hosted clusters at the GPU Test Drive Website.
Gupta told eWEEK that Nvidia’s background and expertise is a key differentiator from AMD, which he said is more focused on its accelerated processing units (APUs) being used in notebook and desktop PCs.
“We don’t see them in big data or HPC announcements,” he said.
AMD officials believe the company is building a strong heritage in HPC, noting that its current FireProS1000—which offers 6GB of GDDR5 memory—was used by the University of Frankfurt’s Institution of Advanced Studies in its SANAM system, which ranks in the top five on the Green500 list of the world’s most energy-efficient supercomputers powered by graphics processors.
The FirePro S10000 12GB Edition will be available in the spring of 2014. The new graphics card supports PCIe 3.0, is optimized for use with the OpenCL programming language and is based on AMD’s Graphics Core Next (GCN) architecture.
AMD officials said the new GPU is good for visualization workloads, such as finance, aeronautics and medicine; double-precisions tasks (including genetic sequencing and computational fluid dynamtics); single-precision (seismic processing, satellite imaging and video enhancement); and ultra high-end workstation jobs (oil and gas and computer-aided engineering).