Intel, clusters on the rise in top 500 supercomputer list

Intel, clusters on the rise in top 500 supercomputer list
PHOENIX — Intel Corp. and cluster architectures are the big winners in the latest list of the world's most powerful computers which was formally announced at the Supercomputing 2003 conference here Monday (Nov. 17).

Intel processors — Xeon, Itanium 2 and a handful of Pentium CPUs — power 189 systems on the list, up from 119 in the rankings released in June and just 56 in last year's list. Many of those Intel-based systems are clusters of commodity computers that now represent the most common architecture in the top 500 at 208 systems up from 149 in June.

"For many years there were no Intel-based systems on the list at all, but that has changed with the rise of clusters," said Erich Strohmaier, a computer scientist with Lawrence Berkeley Laboratories who maintains and helped launch the top 500 in 1993. "IBM has sold more than 70 large [Intel] Xeon clusters and Hewlett-Packard is shifting from its PA-RISC to [Intel] Itanium which also is driving this trend," he added.

Indeed, 152 of the top 500 systems now use 32-bit X86 Intel Xeon CPUs, typically dual-processing versions. Thirty-two Intel Itanium systems made the current list. Among other x86 variants, 13 systems on the list use Advanced Micro Devices' Athlon or Opteron CPUs, including the sixth most powerful system.

While the overall rate of performance growth of the top 500 systems is steady, distributed clusters of generally off-the-shelf microprocessors are taking larger shares of these high-end systems. Indeed, seven of the top 10 systems are now clusters. The exceptions are NEC Corp.'s Earth Simulator, by far the most powerful system in the world at 35.8 Tflops and two IBM SP-class proprietary systems ranked 8th and 9th.

"It's really surprising that clusters are now showing up in such high numbers with such high performance rankings, but there is still a lot of variety in the kinds of cluster processors and interconnects we are seeing," said Strohmaier.

While the vast majority of the clusters are using Linux variants as their operating system, CPUs include a broader variety of IBM, Intel and AMD processors. The Infiniband interconnect is making its first appearance in the top 500 in systems including the Virginia Polytechnic Institute's cluster of Apple G5 systems that ranks number three on the list. Other more traditional interconnects widely represented among clusters include Myranet and Quadrics.

Aggregate performance of systems on the current list is 528 Tflops, up from 375 Tflops in June. The lowest-end system now making the list packs 403.4 Gflops, up from 245.1 Gflops, and to get into the top 100 systems now must have 1.142 Tflops, up from 708 Gflops.

China is on the rise on the top 500 list with an Itanium 2 system built by Legend Computer for the Chinese Academy of Science ranking number 14. Previously, China's most powerful system ranked 51st.

IBM is also raising eyebrows with its dense BlueGene/L prototype, hitting number 73 in a package just half a single rack in height. "It's astonishing to see a system of this physical size get such a high ranking," said Strohmaier.

Al Gara, chief architect for BlueGene/L, said IBM is taking a novel approach in trading off processor performance for scalability in the system. BlueGene/L uses a custom version of a 1W PowerPC 440 with two processor cores and two double floating point units on each chip, giving eight floating point operations per clock. The combination of the muscular but low power CPU and a handful of proprietary chip-to-chip networking techniques will allow designers to scale the system to 64,000 processor chips in a system that uses few electronics beyond the custom CPU and commodity DRAM.

The chip-to-chip interconnects in BlueGene/L include a novel 3D torus net for direct links between processors with adaptive routing in hardware, a tree net for handling simple global arithmetic operations, a barrier network to stop the machine within microseconds for diagnostics and error handling and a "back door" JTAG network used for noninvasive monitoring and booting the system.

IBM still commands the lion's share of total installed performance on the top 500 list at 35.4 percent of total performance (up from 34.9 percent in June), followed by Hewlett-Packard (22.7 percent) and NEC (8.7 percent). In terms of total number of systems on the list, HP still grabs the top spot with 165 systems (up from 159), IBM comes in second with 159 (up from 158) and SGI is third at 41 (down from 54). Cray has 10 systems on the list, it's most powerful ranked number 19.

"It's getting to be a two-player game among systems vendors with IBM and HP and other companies much lower down on the list," said Strohmaier.

The top 500 list is a ranking based on scores on the Linpack benchmark, essentially a set of linear equations widely used in computationally intensive applications. At the SC2003 conference some of the researchers involved in the top 500 ranking will announce a new benchmark dubbed HPCchallenge that will incorporate Linpack and four additional metrics.