Users of high-performance computers are already making plans for x86-based, quad-core processors due from Intel next month and from Advanced Micro Devices by mid-2007.
By year's end, the Louisiana Optical Network Initiative (LONI), a network of high- performance computers run by eight state research universities, expects to have 1,440 Intel quad processors, code-named Clovertown, running on Dell servers and delivering 50 trillion floating-point operations per second (TFLOPS) of power.
Officials declined to disclose the cost of the systems.
When combined with the installed IBM p5 servers running various Intel processors, the capacity of the LONI network is expected to reach 85 TFLOPS.
Meanwhile, the Texas Advanced Computing Center at the University of Texas in Austin intends to buy 13,000 quad-core processors from AMD to run on Sun Microsystems servers. The quad-core Opteron chips are due out by mid-2007, according to AMD.
The US$59 million purchase will be funded by the National Science Foundation as part of its project to build a 400-TFLOPS system.
Widespread use of the quad systems depends on the development of software that can take advantage of the technology, noted John Fruehe, worldwide business development manager for Opteron at AMD. Each core in a multicore processor runs at a lower clock speed than single-core chips, and performance gains depend on how well the software can run in parallel environments, he said. "So much more relies on the software than ever before," Fruehe said.
Boyd Davis, director of marketing for the server products division at Intel, noted that many server-based business applications have already been adapted for multicore environments. "In a server market, most of the applications that customers care about have already been built with multithreading in mind," he said.
Many high-performance applications, such as those the Texas Advanced Computing Center plans to run on its Opteron quad chips, use the Message Passing Interface (MPI) communication protocol to work in large parallel environments. Tommy Minyard, assistant director at the center, said researchers there plan to find ways to improve the protocol's ability to increase the performance of applications running on quad chips.
One problem with the protocol, Minyard said, is that it causes network communications to use significant processor overhead. "If processors are trying to communicate with one another, they are not doing floating-point operations per se," said Minyard, who plans to work on developing hybrid techniques that would limit or move MPI's communications work off the processor. "We want to maximize the amount of floating-point operations that can be done with these systems at large scale."
In an effort to encourage businesses to utilize high- performance computing systems, 10 percent of the computing capacity in the LONI project will be reserved for businesses whose use of the system will create jobs in the state, said Charles McMahon, executive director. He said the group expects that companies willing to come to Louisiana and create jobs can benefit by using some of its compute cycles.
McMahon said LONI members hope to have their quad system in production as early as January.