Welcome to the MacNN Forums.

b3t4 · Aug 30, 2003, 12:57 PM

http://thinksecret.com/news/virginiatech.html

pretty crazy idea.

beta

businezguy · Aug 30, 2003, 01:04 PM

Well, that certainly bumps up the status of Apple's machines. I'm sure we will hear more about this if Apple manages to build the G5 into blade servers in the future, although that would most likely require a .9 micron process before that is made possible.

stuffedmonkey · Aug 30, 2003, 02:04 PM

Heh - as A Virginia Tech alum - I would like to take this chance to say Go Hokies!

Super Glitcher · Aug 30, 2003, 02:14 PM

That's the coolest thing I've ever seen in my life.

What are the chances that they'll let me use it do some rendering in Lightwave?

Shoot.. with that much juice I could even turn on radiosity

docholiday · Aug 30, 2003, 02:14 PM

Amazing..

I have a question though: Does anyone know how the tests are performed? Are they based on a software or how does it work? (just curious)

Ganesha · Aug 30, 2003, 04:21 PM

If true, this could explain the shipping delays.

MindFad · Aug 30, 2003, 04:30 PM

Yeah, but how many FPS will it get in Quake 3?

saru boy · Aug 30, 2003, 04:31 PM

So...did they get those G5s?

:XI: · Aug 30, 2003, 04:35 PM

Originally posted by businezguy:
Well, that certainly bumps up the status of Apple's machines. I'm sure we will hear more about this if Apple manages to build the G5 into blade servers in the future, although that would most likely require a .9 micron process before that is made possible.

1. Wasn't there an announcement that IBM were going to be using the 970 in blade servers?

2. Why would the 970 require a '.9 micron process' to be used in blade servers?

alien · Aug 30, 2003, 05:26 PM

Originally posted by docholiday:
Amazing..

I have a question though: Does anyone know how the tests are performed? Are they based on a software or how does it work? (just curious)

The LINPACK test that Top500 uses is a measure of floating point execution rate. It takes a linear system of equations and solves it for the unknowns. Like this one:

0.84x + 0.52y + 1.23z = 0.45
8.15x + 0.23y + 2.34z = 1.23
0.54x + 1.33y + 3.22z = 2.21

Only of course, in scientific computing, these systems are MUCH bigger.

Here, n=3 (unknowns), but the user can scale the problem to whatever will give the highest score. For the top super computers today, n can be several hundred thousands.

To solve the problem (finding x, y and z), LINPACK divides the job over all the processors and then each does floating point additions and multiplications over and over again (and some other operations like division too), according to an algorithm called LU factorization. Processors also communicate between each other.

LINPACK isn't solving this problem in the most efficient way, mostly due to poor memory access patterns. But the benchmark is recognized as a decent measure of real-world performance in numerical linear algebra, which this is an example of, and which constitutes the bulk of scientific and technical computing.

alien · Aug 30, 2003, 06:46 PM

I'm trying to speculate a little on what to expect of LINPACK performance, and thereby ranking on Top500.

The top-ranking IBM p690 machine based on Power4 has 1280 dual-core chips, each core has 2 fp units, and each running at 1.3GHz. This gives Rpeak = 1280*2*2*1.3G = 6656G.
This is theoretical peak performance.

Rmax, the maximal LINPACK performance actually achieved is about half of that, at 3241G.

So, 1100 G5s each having 2 cpus, each with 2 fp units, and each running at 2.0GHz, should mean that Rpeak = 1100*2*2*2G = 8800G.
If the G5 cluster were to be among the top five, Rmax today would have to be more than 7304G, but this may well be too little when the next list is announced.

Think Secret says they're aiming at more than 10T (10000G). Does that mean they're counting a floating point mult-and-add as 2 flops? (Which the LINPACK doesn't it seems.)

(LINPACK is measuring full precision performance, so AltiVec shouldn't be a factor here.)

I'm a little confused. Looks to be one hell of a supercomputer though.

I wonder how good Infiniband is on this scale. That will mean a lot when it comes to real-world performance solving numerical problems.

Brazuca · Aug 30, 2003, 07:04 PM

Originally posted by alien:
I'm trying to speculate a little on what to expect of LINPACK performance, and thereby ranking on Top500.

The top-ranking IBM p690 machine based on Power4 has 1280 dual-core chips, each core has 2 fp units, and each running at 1.3GHz. This gives Rpeak = 1280*2*2*1.3G = 6656G.
This is theoretical peak performance.

Rmax, the maximal LINPACK performance actually achieved is about half of that, at 3241G.

So, 1100 G5s each having 2 cpus, each with 2 fp units, and each running at 2.0GHz, should mean that Rpeak = 1100*2*2*2G = 8800G.
If the G5 cluster were to be among the top five, Rmax today would have to be more than 7304G, but this may well be too little when the next list is announced.

Think Secret says they're aiming at more than 10T (10000G). Does that mean they're counting a floating point mult-and-add as 2 flops? (Which the LINPACK doesn't it seems.)

(LINPACK is measuring full precision performance, so AltiVec shouldn't be a factor here.)

I'm a little confused. Looks to be one hell of a supercomputer though. I wonder how good Infiniband is on this scale. That will mean a lot when it comes to real-world performance solving numerical problems.

I was under the impression that the G5 was entirely full-precision. I understand that it wasn't but can anyone confirm this?

AssassyN · Aug 30, 2003, 07:27 PM

Originally posted by Ganesha:
If true, this could explain the shipping delays.

Considering over 100,000 G5's were pre-ordered, that's doubtful.

stuffedmonkey · Aug 30, 2003, 11:45 PM

Wierd. I used to work for Virginia Tech Computing services when I was in school. The building they are going to put the "supercomputer" is where I used to ahve to go to turn in my time sheet! Some of the VT computer services guys are huge mac geeks - unfortunatly the campus is full of wintel guys...Oddly both my roommate and I both graduated from there with liberal arts degrees, and ended up fixing macs for a living - he does freelance in NYC - I ended up working for Apple for two years as a mac genius.....

chrisutley · Aug 31, 2003, 03:01 AM

I wonder if the Macs will be running OS X or Linux?

Didn't NASA just buy a bunch of Xserve boxes that they are installing Linux on?

Angus_D · Aug 31, 2003, 05:27 AM

Originally posted by chrisutley:
I wonder if the Macs will be running OS X or Linux?

Didn't NASA just buy a bunch of Xserve boxes that they are installing Linux on?

They did, but I doubt Linux runs on the G5s just yet. Probably a few more months.

alien · Aug 31, 2003, 05:32 AM

Originally posted by Brazuca:
I was under the impression that the G5 was entirely full-precision. I understand that it wasn't but can anyone confirm this?

It does have 2 powerful full-precision (64-bit) floating point units that each does multiplication and addition.

Furthermore it has the AltiVec unit, which works on 4 32-bit floating point numbers simultaneously. But I don't think this can be used in this benchmark.