Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Originally posted by srobert
Hell! their list is oudated! They don't even list the Virginia Tech Cluster

The list is updated twice a year (Spring and Fall). One of the reasons for the rush and hard deadline was the opportunity to make the Fall Top500. Had they waited until Spring, VA Tech would have gone against a number of larger Opteron and Itanium clusters using Infiniband and might not have held such a prominent position. For instance, it would still have been the cheapest cluster on the list, but not the 10x cost savings--more like 2-3x. Also, if you networked more Opterons or Itaniums (a lot more in the case of Itaniums--when the Fall Top 500 comes out, look at #4 and extrapolate) you could knock Terascale from the #3 slot which is all but assured for the Fall list (it's currently 1 Tflop higher than the #4 and no new machines are eligible for the Fall list).

Take care,

terry
 
Use of super computers. Surprisingly many things gets very very complicated, when you start to calculate carefully. Not only weather and such things but also things like how a car engine runs, airflow around a house.

Just calculate how a stream of milk mixes with the hot coffe. Hot coffe has less density than cold, add suger and the density goes up. turbulent liquid flows made even more complex by the temparature gradients and so on. More or less endless variables. Not that many would like to spend a lot of money on just that but the point is that for careful calculations almost anything become really lots and lots of numbers.

Many things can be broken down to distributed computing but I assume that not all kinds of calculations work that way. Iterative work should be hard to run distributed.

The G4 started as "the first personal supercomputer" got stuck for ages at 500 MHz and then have been beaten by AMD and Intel in 2000 and 2001 and 2002 and 2003😡
The G5 seem to be of to a better start, we will get some kind of long term indications next year😀
 
Check the numbers - VTech cost is going up

Originally posted by tychay
For instance, it would still have been the cheapest cluster on the list, but not the 10x cost savings--more like 2-3x.

It's hard to meaningfully compare the costs of the VTech cluster against the others.

The VTech $5.2M figure doesn't include facilities, air-conditioning, power, installation, staffing, ....

The Earth Simulator figure includes the building and seismic reinforcements.

The 10x savings number rapidly drops when comparable methods for calculating the expenses are used!


Had they waited until Spring, VA Tech would have gone against a number of larger Opteron and Itanium clusters using Infiniband and might not have held such a prominent position.

A good point to emphasize - the VTech cluster is the first large scale InfiniBand fabric cluster on the list. The Itanium, Xeon and Alpha clusters are using less capable fabrics.

When other clusters using InfiniBand show up, we'll see that the fabric is an important factor in the VTech cluster's performance - it's not all due to using PPC970 systems....

And wait for IBM BladeCenter systems with IB and the PPC970 - you could put 2200 PPC970 CPUs in a standard 10ft x 10ft (100 sq ft) cubicle !! Compare that to the 3000 sq ft VTech cluster....
 
sour grapes - the point is by going with Apple, VA Tech was able to get their system online today. Going the other routes, they would have missed their deadline. So what if there will be other clusters online next year that are faster than what they have now - that will ALWAYS be the case. There will be 3 GHz G5 clusters next year too - those will be 50% faster as well - that'll be competing with the new itaniam, Xeno and alpha clusters. And don't for a second overlook the possibility that some planned clusters switch over to G5 based because of the Va Tech result. Don't think the people designing those clusters in the works aren't getting asked questions about why they are spending more money for less performance. The system Va Tech built is competing with and kicking the butts of the existing Alphas an itanium systems, several of which are brand new systems as well.

Just look at #3 vs. #4. Both are coming online today, both are a comparable number of CPUs, #3 is far cheaper no matter how you look at it. Yet, the Rmax for the G5 machine is coming darn close to the Rpeak for the #4 system. the Rmax has now passed the Rpeak for brand new alpha systems with about half of the CPUs in the alpha system.

This is all with stock desktop systems not even designed with supercomputers in mind. Now that this has happened, you can be sure Apple is designing boxes for the supercomputer market.

Aiden - your arguments have been pretty weak - and are getting weaker with each new performance report.
 
Re: Yikes

Originally posted by itsbetteronamac
I find it kinda scary that a weapons company is looking into somthing with that much proceessing power. I mean think of what low-cost Supercomputers in the wrong hands can do.

Just my thoughts
I'm not pro-war, but I would like our armed forces to have the best possible weapons/defense should they need to protect themselves or private citizens. These companies are going to find the computing power they need from someone, be it Dell, IBM, or HP; so why not use an Apple/Mac based solution? If helps them with the design of the next generation armor or bullet or whatever, so be it.
 
Re: Check the numbers - VTech cost is going up

Originally posted by AidenShaw
It's hard to meaningfully compare the costs of the VTech cluster against the others.

The VTech $5.2M figure doesn't include facilities, air-conditioning, power, installation, staffing, ....


you are wrong there! ...it didn't include a building ...but it did include all of the cooling/power equipment and the installation cost! (i believe the installation cost them something like 600 pizzas)
 
Originally posted by freemidnight
I imagine Crae, HP, Next and the rest buying in secret G5 and saying that their cluster is one of the fastest on the planet!😀 😀 😀

It's Cray, and NeXT was bought by Apple in 1997. You know that OS X thing? Ya, roots from them.
 
Re: Check the numbers - VTech cost is going up

Originally posted by AidenShaw
A good point to emphasize - the VTech cluster is the first large scale InfiniBand fabric cluster on the list. The Itanium, Xeon and Alpha clusters are using less capable fabrics.

When other clusters using InfiniBand show up, we'll see that the fabric is an important factor in the VTech cluster's performance - it's not all due to using PPC970 systems....

Actually the Itanium systems have the same theoretical Gflop rating as the 970 at the same CPU speed, but since the Itanium is only available at 1.5 Ghz its top end is only 3/4 that of the 970. I would assume the Opteron @ 2Ghz has the same theoretical rating (but I haven't seen it listed, so I don't know) but since it integrates a memory controller it may have better memory latency than the 970 - but given that the 970 can still fetch more 64-bit floats that it can process per clock cycle (assuming good locality of reference) then I would say it would be a dead heat.

And that is not even including Altivec, which for some types of problems (e.g. encryption/compression research) would make for astounding performance increases.

And wait for IBM BladeCenter systems with IB and the PPC970 - you could put 2200 PPC970 CPUs in a standard 10ft x 10ft (100 sq ft) cubicle !! Compare that to the 3000 sq ft VTech cluster....

Now that would be sweet! Of course it would probably have to use liquid nitrogen for a cooling system 😀 .
 
Re: Re: Check the numbers - VTech cost is going up

Originally posted by zweigand
you are wrong there! ...it didn't include a building ...but it did include all of the cooling/power equipment and the installation cost! (i believe the installation cost them something like 600 pizzas)

According to the story on the OS X conference at MacCentral

The system costs $5.2 million for the G5s, racks, cables, and Infiniband cards. Virginia Tech spent and addition $2 million on facilities, $1 million for the air conditioning system and $1 million for the UPS and generator.


The point is simply that the costs quoted for various Top500 systems are calculated using different methods, and included different items.

To say that VTech was X times cheaper than another supercomputer is a dubious claim.

Perhaps if "pennies per FLOP" were part of the the Top500 submission we'd see more rigour in specifying the cost, and comparisons would be more meaningful. (The official "cost" for VTech might need to include an estimate for the market value of the volunteer labour and pro-rated expenses for the space in the building, for example.)

It was a cheap cluster, but try to avoid the "X times cheaper" comments when the figures aren't comparable due to omissions and special circumstances.
 
Originally posted by JohnGillilan
This may sound stupid, but what is a "supercomputer" even used for?? Is it partially just for bragging rights, or do these clusters have practical purposes??

Testing nuclear weapons without actually making Pacific atolls uninhabitable. (This application is a big reason why export of supercomputers is restricted).
 
Re: Check the numbers - VTech cost is going up

Originally posted by AidenShaw
It's hard to meaningfully compare the costs of the VTech cluster against the others.

The VTech $5.2M figure doesn't include facilities, air-conditioning, power, installation, staffing, ....

The Earth Simulator figure includes the building and seismic reinforcements.

The 10x savings number rapidly drops when comparable methods for calculating the expenses are used!

Unlike you, I know something called "Fermi" estimation. The $5.2 M figure included the cost of the networking and cabling which in Q&A was determined to be $1.5 million putting the cost at $3.7 million which is the educational price + 2GB RAM (no discounts). This is significantly below the $9-10 million quoted price for the Itanium systems by a factor of 2-3.

Now add in the cost of the facilities upgrade and air conditioning ($2 million according to the presentation), chuck in a couple of million for the building and you'll see a more than 10x price differential estimation vs. the $150 million (estimated) #2 or $300 million (estimated) #3. If anything, it's probably greater than 10x (of course, it's a bit unfair here because I it is impossible for Terascale to be #1 and it's highly unlikely it will beat #2).

I seem to remember the new #4 cluster weighs in the at $30-60 million (i.e. Terascale only offers a 2-3x cost savings, I figure it is closer to 2x due to the chips and 3x because they used Infiniband).

A good point to emphasize - the VTech cluster is the first large scale InfiniBand fabric cluster on the list. The Itanium, Xeon and Alpha clusters are using less capable fabrics.

I'll take exception to the wording "less capable." I believe the fabrics are just as capable, just not close to affordable for a university. $1.5 million for cards, routers and cabling: that's truly bargain basement. (Thanks, Intel!)

However the general idea of your statment bears repeating 100x over because I believe this is why the rush occurred which forced them to by G5 systems instead of wait until the 4-way 970's start shipping from IBM in 1Q 2004. Still the 970 is the clear winner on price/performance as by it's 2x cheaper quoted price even with such useless parts as a superdrive, graphics card, optical in/out, usb 2, firewire....

BTW, it's somewhat less than 50% better than the Itanium (the Rpeak is 50% greater because they have the same flops/clock but the best Itanium clocks in at 1.5Ghz. Note though that the Itanium is more efficient in other parts so I add the qualification... comparing to the #4 cluster it seems it's between 25-50% better) and it's 100% better than the Opteron (same Mhz ratting but half the flops/clock). This is my rationalization of the above statement. I don't know the costs but they're all in the same price range in my little "rough estimation" book of "same process = same costs".

And wait for IBM BladeCenter systems with IB and the PPC970 - you could put 2200 PPC970 CPUs in a standard 10ft x 10ft (100 sq ft) cubicle !! Compare that to the 3000 sq ft VTech cluster....

I don't think these will be able to be switched using infiniband (yet). But there are 4 ways coming down the pike that are in standard rackmount configs.

Personally, I can't wait for the 970 blades. We have a number of empty slots in our BladeCenter at work.
 
individual G5 performance

VaTech seems to have squeezed 4.67272727... Giga Flops out of each G5 processor, meaning that they've achived a 9.35 Gflop mark for each machine... The #1 ranked Earth Simulator managed 7.004 per processor, while #2 ASCI only could pull off 1.7 (they had over 8000 processors) Of intrest, HP's cluster with 1.5 Gig Itanium's got 4.459.

Numbers calculated by dividing (Rmax GFlops) by (# Processors).

Historical Note: Supercomputers (defined as a cpomuter with sustained performance of over 1 Gflop) were once banned from exportation because of their potential application in weapons desing and/or production. The G4 was declared the worlds first personal supercomputer based on the above definition...
 
another G5 Dualie... performance on this one suffers

interestingly enough, the only other G5 (dual 2Ghz with infiniband 4X Fabric) delivered disappointing performance. a per processor performance of 3.21 Gflops was 'achieved', using the same methods as before, equating to a measly (in comparison) 6.41 macine performance.

I am not sure if AltiVec is being used, but it would be quite interesting to see it it was. I am sure that VaTech would be well above the current #2 if it wasn't accounted for in the PDF, but was in the 'top 500' rankings.
 
Re: another G5 Dualie... performance on this one suffers

Originally posted by ibilly
interestingly enough, the only other G5 (dual 2Ghz with infiniband 4X Fabric) delivered disappointing performance. a per processor performance of 3.21 Gflops was 'achieved', using the same methods as before, equating to a measly (in comparison) 6.41 macine performance.

I am not sure if AltiVec is being used, but it would be quite interesting to see it it was. I am sure that VaTech would be well above the current #2 if it wasn't accounted for in the PDF, but was in the 'top 500' rankings.

AltiVec is useless for LinPack numbers because it has to be able to do 64pit precision which AltiVec units cannot. End of story.

(As for the G4 being the "first personal supercomputer"-- it was all marketing. There were other machines out on the market at the time that were also banned for export, however Apple decided to use this angle for marketing. If you learn all your info from advertising, you'll be seriously mis-informed)
 
Re: Yikes

Originally posted by itsbetteronamac
I find it kinda scary that a weapons company is looking into somthing with that much proceessing power. I mean think of what low-cost Supercomputers in the wrong hands can do.

Just my thoughts
The person who posted "Argonne Weapons.." didn't know what they were talking about.

Argonne National Laboratories is a research campus in the suburbs of Chicago that is run buy The University of Chicago for the DOE. It isn't a weapons plant, it's a Physics research lab.

You can pick up the shuttle to Argonne from the right in front of the squash courts at UC (well, the Enrico Fermi building across from where THE squash courts used to be.
🙂
 
Re: Re: another G5 Dualie... performance on this one suffers

Originally posted by legion
(As for the G4 being the "first personal supercomputer"-- it was all marketing. There were other machines out on the market at the time that were also banned for export, however Apple decided to use this angle for marketing. If you learn all your info from advertising, you'll be seriously mis-informed)

Well, it was marketing, but it was also the letter of the law.
Congress determined that computers capable of 1GFlop sustained performance were "SuperComputers". OK, it was outdated, but legally the G4 WAS a supercomputer when it was released.

The new standard is just over 50GFlops, so we probably have to wait for the G6 to run into export problems again. ;-)
 
Re: another G5 Dualie... performance on this one suffers

Originally posted by ibilly
interestingly enough, the only other G5 (dual 2Ghz with infiniband 4X Fabric) delivered disappointing performance. a per processor performance of 3.21 Gflops was 'achieved', using the same methods as before, equating to a measly (in comparison) 6.41 macine performance.

I remember seeing that benchmark. It is a fraction of Terascale that was turned online in an early dry run. In other words, it's the same computer just at an earlier stage.

VaTech seems to have squeezed 4.67272727... Giga Flops out of each G5 processor, meaning that they've achived a 9.35 Gflop mark for each machine...Of intrest, HP's cluster with 1.5 Gig Itanium's got 4.459.

The Itanium performs near parity (though decidedly less) than the 970 (G5) because it is an incredibly efficient chip/Mhz (as noted by the manager of Terascale). However, as he also noted, "Ironically, it appears Intel lost the Mhz race." Meaning that it was cheaper (by a lot) to buy a 2Ghz G5 vs. a 1.5Ghz Itanium: price quoted was 2-3x greater and that was the system price. The Apple quoted price was list price and included a whole lot of unneeded extras.

The other reason is Infiniband is new as is experience with the 970 compiler optimizations (Professor Goto's tweaks), as are the kernel patches. You may notice the numbers reported by #3 G5 cluster continue to inch up, but the same isn't true with the #4 Itanium cluster.

Heck I'll even entertain the possibility that Infiniband isn't all that and it'll never break 60% efficiency under that load. We just don't know what the bottlenecks are until more Infiniband clusters come on line--wait until Spring 2004 or Fall 2004 Top500 before the verdict is in.

I suppose the Opteron is probably very efficient--at least, I'd hope so because it's theoretical max is half of the G5. It's a pity because my ex-roommate from grad school does HPC and TPC optimizations for Advanced Micro Devices.
 
One point to remember is the VA cluster is the first supercomputer to be ordered online!

The main man did have to go to the Applestore and order 1100 G5s, it may even be the diggest online transaction sofar on the web?

Viv
 
Originally posted by Viv
One point to remember is the VA cluster is the first supercomputer to be ordered online!

The main man did have to go to the Applestore and order 1100 G5s, it may even be the diggest online transaction sofar on the web?

Viv

I doubt it. There are a lot of B2B markets that we never see that conduct transactions like this often.
 
SuperComputer Kit - Cluster of 12?

Each of 92 racks at the VTech "Big Mac" SuperComputer Cluster holds 12 G5s, for a total of 1,100 G5 dual 2GHz boxes

One rack of 12 @ $3,000 ea. = $36,000
1,100 G5 x $3,000 = $3,300,000 Basic, Off-the-shelf from online Apple Store
but, did they add Fibre Channel PCI card $500, or 2 GB more RAM $1,300 ?
and, from this did they get a 10% education institutional discount?
Building a mini-cluster of just one rack of 12, add $2,000 for the Rack and wiring, and another $2,000 for backup, connectors, some software, $6,000 for 3 Apple 23" Studio Displays and a cool looking editing table & Panther black leather chair (not cat fur).
So, how much for the Kit?
The VTech budget was $5.2-million spread over 5 years, plus they had to build a special facility, air conditioning system and had to tie up to a major internet backbone.

Still, a 12 unit cluster is about 1.09% of their 10.28 TFlop SuperComputer (with a goal to crank it up as high as 13 TFlops), making a per rack cost of about $56,727 related to the whole budget.

So, who would love one rack out of 92 on a $57,000 budget? 😕
Think we could crank out some real time rendering with Final Cut Pro 4.0?

By the end of the year, expect SuperDrives (not necessarily through Apple) to be available in the U.S. at 8x speed, for under $500 (closer to $350), and
16x and even 32x DVD speed may not be that far off after that.

Audio - how about a Digi 001 control board and some studio reference, self powered surround sound speakers. Oh, and an 18" 1,200 watt subwoofer.

What a sweet editing suite that would be. 😛

But not for long, 'cause rumours are speading about 30" Apple HiDef LCD Studio Displays with smaller Aluminum metal frames ($6,000 ea.) - imagine being surrounded by 3 of them mounted close together in a super wide-screen SurroundVideo mode satisfying even peripheral vision.
Flight sims for virtual reality with feedback controls. Want to fly the Concord?

If only there was a real Santa Clause, I'd write him a letter.
All I want for Christmas is a single-rack of 12, SuperComputer G5 mini-cluster running at a mere 0.11205 TFlop.
Hey, Steve, how about an educational discount?
 
Originally posted by JohnGillilan
This may sound stupid, but what is a "supercomputer" even used for?? Is it partially just for bragging rights, or do these clusters have practical purposes??

Sequencing genes.
Simulating nuclear explosions
Surfing for porn. 😀
 
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.