Posted on 06/21/2004 8:56:58 PM PDT by Theo
Apple Computer Inc. will announce on Monday the sale of 1566 dual processor 1U rack-mount 64-bit Xserve G5 servers to COLSA Corp., which will be used to build what is expected to be one of the fastest supercomputers in the world. The US$5.8 million cluster will be used to model the complex aero-thermodynamics of hypersonic flight for the U.S. Army.
"We did about a year and a half of research on a variety of processors before making our decision," Dr Anthony DiRienzo, executive vice president at COLSA Corp., told MacCentral. "We did a best value competition and Apple won that competition. It was based on performance; the facility (power requirements, floor space etc.); cost; and an assessment of vendor stability. We solicited to six companies and they won."
The supercomputer, named MACH 5, is expected to deliver peak performance capability of more than 25 TFlops/second. In comparison, the Virginia Tech supercomputer announced last year attained sustained performance of approximately 10 TFlops/second, according to Apple director of product management, server hardware, Alex Grossman.
With those numbers, the MACH 5 would rank second only to Japan's $350 million Earth Simulator computer.
"We evaluated PC-based proposals from other vendors but none came close to delivering either the price, performance or manageability of the AppleXserve G5," said DiRienzo.
The Xserve G5 supercluster system is expected to be online and working for the Aviation and Missile Research, Development and Engineering Center (AMRDEC) division of the US Army Research and Development Command by late Fall.
Shortly, DiRienzo said they would take delivery of 300 Xserves a day, set them up in the racks and the next day begin the process again until all of the Xserves are installed and working.
"We expect MACH 5 to rank as one of the most powerful supercomputers on the planet," said Dr Anthony DiRienzo, executive vice president at COLSA Corporation. "According to the November 2003 Top 500 supercomputer list, it would rank second only to Japan's $350 million Earth Simulator computer at less than two percent of the cost. We evaluated PC-based proposals from other vendors but none came close to delivering either the price, performance or manageability of the Apple Xserve G5."
Amazing times we live in, where it's come to the point where almost anyone can own a staggering supercomputer....
Imagine a Beowolf cluster of these. (c) Slashdot
Watch for the IBM "Blue Gene" series -- it is going to demolish everything... there are already two prototype clusters to test things out and the final machine will be 10x faster than the Japanese Earth Simulator, which has had a long reign as #1...
If "Blue Gene" weren't enough, just wait for "Red Storm". It is going to be an awesome machine based on 8 to 16 thousand Opterons -- and is a fair bit cheaper per Teraflop than the Apple offering...
""We expect MACH 5 to rank as one of the most powerful supercomputers on the planet," said Dr Anthony DiRienzo, executive vice president at COLSA Corporation. "According to the November 2003 Top 500 supercomputer list, it would rank second only to Japan's $350 million Earth Simulator computer at less than two percent of the cost. We evaluated PC-based proposals from other vendors but none came close to delivering either the price, performance or manageability of the Apple Xserve G5."
If you want to be included or excluded from the Mac Ping list, let me know in Freepmail.
Ft. Meade must be hogging them all, or the Air Force is only letting the Army have rotary-wing computers ;-)
Not a bad choice for CFD, since it is completely compute bound and is one of those narrow codes that PPC970 would excel at.
Would that be the VT one that is currently having it's machines swapped out for NEW G5 Xserves? Becasue we all know that the #3 machine from last year just exploded or something becasue it had the name Apple associated with it. Sheeezzzz. Where do these people come from. Do a little research next time before you slam.
BTW, the Xserve takes care of the memory issue to which you are referring. So it's only onward and upward form hear on out as the US Army so nicely points out - "We evaluated PC-based proposals from other vendors but none came close to delivering either the price, performance or manageability of the Apple Xserve G5."
Whoof! That's some schedule. :)
That said, Macs and Windows-based machines seem to be leapfrogging over each other. One year Macs are nasty fast and cheap, the next year windows-based PCs are nasty fast and cheap.
I'm pretty platform-agnostic, although I prefer the Mac platform, and do my software development on a Mac -- I find it makes me more efficient with my time and my time is more pleasantly spent on a Mac....
300 Xserves a day! LOL! :-)
Looking at the June 20th list of the Top 500, had the VT cluster been running it still would have been 4th at a sustained 10.3 Gflops.
I believe it is down because they are re-configuring it from G5 Towers to G5 X-serves. Rumor has it they were waiting for the 2.5 GHz G5s to be available in sufficient quantities in the G5 X serve configuration.
Looking at the expected performance rating of the new cluster, it looks like the Army will be getting the 2.5s.
With that schedule, do you think they would notice of they only got 298 and a couple fell off the truck at my house???
IIRC the G5 can do 4 floating point operations per cycle, so 1566 Xserves * 2 processors/Xserve * 2.0 GHz/processor * 4 flops/cycle = 25.056 Tflops.
One for me! One for me!
Using your formula, the Virginia Tech 2.0GHz system should have
1100 G5 Desktops * 2 processors/G5 Desktop * 2.0 GHz/processor * 4 flops/cycle = 17.6 Tflops.
However, the reported sustained performance was only 10.3 Tflops.
Evidently, some of the Tflops are used in housekeeping routines. Apparently, the true efficiency implied by the formula you used would require a fudge factor of 58.5% of theoretical performance to arrive at a realistic performance rating, thusly:
1100 G5 Desktops * 2 processors/G5 Desktop * 2.0 GHz/processor * 4 flops/cycle * 58.5% = ~ 10.3 Tflops.
We can assume that the same NUMBER of clock cycles are involved doing the housekeeping so the same percentage would not apply to the faster processor. Therefore, the housekeeping would take 17.6 Tflops - 10.3 Tflops = 7.3 Tflops. Let's assume that is a constant: No matter how fast the supercomputer, housekeeping will take 7.3 Tflops. The faster processor would have a greater percentage of its cycles available for actual computing.
Given the above, the theoretical maximum of the cluster claimed would have to be 25 + 7.3 = 32.3 Tflops which would produce an efficiency fudge factor of ~77.4%. Let's see what theoretical performance a 2.5 GHz cluster with 1566 Xserves should produce.
1566 Xserves * 2 processors/Xserve * 2.5 GHz/processor * 4 flops/cycle = 31.320 Tflops.
Multiply times our housekeeping fudge factor percentage:
31.320 * ~77.4% = ~24.242 Tflops!
DANG! That looks like the right ball park!
Working backwards, applying a housekeeping fudge factor to get the 25 Tflops performance from that number of processors, the clock speed of the processors is about 2.5GHz
bttt
The VT machines were already taken down, and sold on eBay!
The latest Top 500 Supercomputer list was just released. Check it out:
http://www.top500.org/lists/2004/06/
The average life span for a Linux cluster is only about 18 months.
A Top 500 Mac was only possible since the G5, and VA Tech was the first. IIRC, general release of G5s was delayed a but while VA Tech took delivery of 1,100 of the early ones. I don't know why they're not on the new list since their old score was enough to keep them at #5 this time, getting nudged down by two entries, LLNL and a BlueGene. It could be they took it down to replace with G5 cluster servers as they were planning.
At a recent High Performance Cluster conference in Austin, the Apple boxes got low marks in how the memory is connected up -- a serious drawback which is why they are not very popular as are Xeons, Opterons, Itanium2 or even IBM's PowerPC blades.
As opposed to Opterons (same class) and Itaniums (different class) I can understand, but the memory performance in Xeons is abysmal in comparison to the G5.
Watch for the IBM "Blue Gene" series -- it is going to demolish everything..
Just wait for a cluster of Opteron-based Cray XD1s. But the point to the G5 clusters is that they're insanely cheap. Notice this one only cost slightly more than the 1,100 node one at VA Tech -- now they're probably using the stripped-down 1U cluster servers instead of full desktop machines.
Disclaimer: Opinions posted on Free Republic are those of the individual posters and do not necessarily represent the opinion of Free Republic or its management. All materials posted herein are protected by copyright law and the exemption for fair use of copyrighted works.