IBM on Tuesday updated its System x iDataPlex servers to include graphics processors from Nvidia as it tries bring high-performance capabilities to its scale-out servers.
The new servers are aimed at organizations that want to run scientific and commercial applications on graphics processors, which are faster at executing certain tasks than traditional CPUs, said David Turek, vice president of deep computing at IBM’s Systems and Technology Group.
The organizations could include scientific labs run by governments or oil and gas exploration companies, where software is mostly developed internally from scratch, Turek said. By tuning elements of an application for purpose-built processors, it’s possible to generate much faster execution speeds than with CPUs.
There is increased interest in hybrid computing systems that use graphics processors along with CPUs, with many companies preparing to deploy such systems, Turek said. Some of IBM’s clients have been testing hybrid servers combining CPUs and GPUs over the last two years and are ready to put them into production, he said.
The iDataPlex dx360 M3 two-socket server will use Intel’s Xeon processors and include Nvidia’s Tesla M2050 GPUs.
The first iDataPlex servers were launched in 2008 and were pitched as an alternative to rack-mount servers for large, scale-out computing environments. IBM at the time said iDataPlex used up to 40 percent less energy while boosting computing power by up five times. Hewlett-Packard and Dell also sell such “hyperscale” computing systems.
iDataPlex was originally pitched as a foundation for cloud computing, said Charles King, principal analyst at Pund-IT. Adding graphics processors further differentiates iDataPlex, which is already one of the highest-performing blade-style systems out there, he said.
“This is evidence that IBM is seeing that architecture as being more flexible than a foundation for cloud computing,” King said.
IBM in 2008 developed a hybrid Roadrunner supercomputer for Los Alamos National Laboratory. The hardware had combined dual-core AMD Opteron CPUs and IBM Cell processors, where algorithms were written in a way to break up computing jobs between the processors. The supercomputer provides more than a petaflop of performance, and was ranked the second fastest supercomputer according to the Top500 server list issued in November 2009 by Top500.org.
“What we’re seeing here is a similar sort of division of labor,” King said.
IBM did not detail performance improvements. However, Nvidia said that iDataPlex dx360 M3 systems with its GPU improved performance by up to 10 times compared to previous iDataPlex servers.
IBM chose Nvidia’s GPUs for its servers because of expanding use of CUDA, a set of programming tools from Nvidia to develop and manage parallel task execution, Turek said. Another factor is the increased adoption of OpenCL as a standard to write applications that can be executed on GPUs. Companies including Intel, Apple, Advanced Micro Devices and Nvidia also back OpenCL.
IBM did not immediately provide information about pricing or availability for the server.