When we run benchmarks on desktop systems, especially Macs, we tend to use standard benchmarks that relate to typical tasks: Photoshop, rendering, Quake, and so on. However, supercomputers are benchmarked on how fast they can run floating point computations in a parallel computing network with tens of thousands of nodes. On the same basis, how do those computers compare to each other?
The worldis fastest computer computers are benchmarked every year at the supercomputing conference. Supercomputing 08 "SC08" is currently under way in Austin, Texas and brings together the worldis experts from government and industry on supercomputing technologies.
"Roadrunner" at Los Alamos National Laboratory, NM (2008)
These computers are benchmarked with software called "Linpack" which assesses how fast a computer can solve a system of linear equations. The metric is the speed of the computer in floating point operations per second, FLOPS.
In late 2008, two computers were benchmarked at over a petaflop, (quadrillion floating points operations per second) "Roadrunner" at Los Alamos National Laboratory in New Mexico, and "Jaguar" at the Oak Ridge National Laboratory in Tennessee. They achieved 1.105 and 1.059 petaflops respectively. They do that by connecting a large number of discrete computers, called nodes, amounting to hundreds of thousands of cores, with a high speed network, called Infiniband. Linux is the favored operating system because it can be tailored to suit the needs of computer scientists and has developed a mature interface to Infiniband.
To make the "Top500" list, the fastest 500 computers in the world, a system must be capable of at least 12.6 teraflops (trillion floating point operations per second).
Comparison to Mac Pro
How would the fastest Macintosh compare? Letis take a Mac Pro running at, say, 3.2 GHz and 8 cores. Each core of a Xeon has two floating point processors and can actually complete two floating point operations for each "tick" of the system clock. (It does that with a neat trick that combines a multiply and an add operation in one clock cycle.).
Apple Mac Pro
As a result, our Mac pro can theoretically achieve:
3.2 GHz * 2 FPUs * 2 FLOPs per cycle * 8 cores = 102 gigaflops.
In practice, itis possible to achieve an efficiency of about 89 percent with Linpack on a Mac Pro, according to Dr. Srinidhi Varadarajan at Virginia Tech University. So the net computational power of the Mac Pro is 91 gigaflops using that benchmark.
Thatis a ratio of 12,142 to 1 in speed for Roadrunner at Los Alamos. Roadrunner, by the way, cost US$120M, about 27,270 times as much as a $4,400 Mac Pro, so weire actually getting a computational bargain, at least in terms of gigaflops per dollar.
The CRAY-1 at Los Alamos (circa 1976)
Hereis a thought. The very first Cray, the CRAY-1, installed in 1976 at Los Alamos was only capable of 100 megaflops. A Mac Pro is about a thousand times faster.
Apple no longer participates in the annual Supercomputing conference, as they did ever since the year 2000. The company has taken a different, more consumer oriented direction -- and probably has made a lot more money in the process.