Who knew the video demands of computer games like Halo would also lead to advances in high performance computing (HPC)? NVIDIA did, and the results of that insight were announced May 15th at the GPU Technology Conference in San Jose, Calif.
Based on its gaming-oriented Kepler architecture yet designed for HPC environments, NVIDIA’s new K10 and K20 Tesla GPUs include 192 cores and support 32 message passing interface queues, versus the one queue supported by the company’s earlier Fermi GPU architecture. The improvements enable systems to process vastly more parallel instructions with far lower CPU utilization.
“We've packed more than seven billion transistors on a single die, giving three times the performance per watt,” says Roy Kim, senior product manager at NVIDIA. “One K10 GPU provides 4.85 teraflops single-precision floating point operations, and 320 gigabytes per second of memory throughput.”
While HPC is only one of four product areas for NVIDIA, along with general purpose computing, gaming, and mobile devices, Kim considers it a small but important market. In fact, the company believes that up to 60 percent of all HPC applications will be modified to support GPUs by the end of 2012.
NVIDIA sells all of its HPC products through the channel, either directly or through OEMs. “SuperMicro sells our products in their systems to hundreds of VARs,” says Kim. “Kepler products are being put more and more into mainstream servers. Dell includes our GPU in their highest volume server. Silicon Graphics also announced a new system using the Tesla K10.”
Both the Tesla K10, which is available now, and the Tesla K20, which will be available by the end of the year, are provided on dual-slot PCIExpress boards. Other vendors coordinating with NVIDIA for the Tesla launch include HP, IBM, and Appro International. Most modern motherboards can support a GPU, and Kim expects OEMs to build systems around different GPU models for differentiation. Tesla K10 GPUs are plug-compatible with existing NVIDIA GPUs for easy upgrading via hardware card swap.
Common applications for GPU-driven systems are financial applications, law enforcement forensics systems, simulation solutions, defense-related signal and imaging processing, and seismography solutions for oil and gas businesses. One early adopter of the Tesla K10, Brazilian energy company Petrobas, reports their seismic application now runs 1.8 times faster yet uses the same amount of power, NVIDIA says.
Universities are another target market for the new chips. “With our Kepler architecture and Tesla K10 and K20, every university has the rack space and budget to deploy their own petaflop supercomputer,” Kim says. The new products could even turn research computing departments into profit centers, NVIDIA believes, as current supercomputers are often oversubscribed with waiting lists.
NVIDIA has yet to announce pricing for its latest GPUs, but Kim expects price points for the Tesla K10 to be similar to NVIDIA’s earlier Fermi GPUs, which sold for anywhere from hundreds to thousands of dollars each, depending on configuration.