Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Virginia Tech Supercomputer Cluster Photos
- Thread starter MacRumors
- Start date
- Sort by reaction score
Probably not until they actually have it up and running. But we can estimate; they have 2200 G5s at 2 GHz each. Even if overhead is such that each processor only averages 1 operation per second of "real" work, that's 4.4 teraflops. If they can take advantage of Altivec, that number will be much higher.
It's good PR for Apple that a supercomputer can be built from G5s and used for such high-end research applications.
It's even BETTER PR that doing so offered more power for less money than all other hardware and OS options, including Dell and including Linux PPC--and as VA Tech announced, that's exactly the case, even though Dell offered special pricing.
But it's better PR STILL that VA Tech had such a tight deadline, and managed to build a homemade 1100-node supercomputer in a few short weeks!
(I wonder what vendor is providing the secondary storage? XServe RAIDs or something else?)
It's even BETTER PR that doing so offered more power for less money than all other hardware and OS options, including Dell and including Linux PPC--and as VA Tech announced, that's exactly the case, even though Dell offered special pricing.
But it's better PR STILL that VA Tech had such a tight deadline, and managed to build a homemade 1100-node supercomputer in a few short weeks!
(I wonder what vendor is providing the secondary storage? XServe RAIDs or something else?)
VT's internet connection may be suffering from the hurricane, cable internet has been off all day in Blacksburg (town VT is located) and not to mention all of us mac lovers that have bombarded that site trying to see pictures.
Like I have posted before, those machines do have video cards in them bc they are most likely going to be resold to students. All of the boxes are being stored to repackage the units in when they are replaced. Unfortunately I never got to assist in the assembly because unboxing and installing the infiniband cards went extremely fast (296 units processed in 2 1/2 hours).
But hey, I still get one of those sweet tshirts for volunteering.
Like I have posted before, those machines do have video cards in them bc they are most likely going to be resold to students. All of the boxes are being stored to repackage the units in when they are replaced. Unfortunately I never got to assist in the assembly because unboxing and installing the infiniband cards went extremely fast (296 units processed in 2 1/2 hours).
But hey, I still get one of those sweet tshirts for volunteering.
They ought to lack optical drives and have the most bare-bones graphics card if any (one you can't BTO officially). They ought to have shipped at a discount without keyboards and mice, too.
But my guess is, VT wanted these NOW--they had an Oct. 1 deadline, and Apple was still the best power for the money. So they accepted standard configs rather than wait. Those extras are a waste in a sense, but I bet it was a deadline issue.
But my guess is, VT wanted these NOW--they had an Oct. 1 deadline, and Apple was still the best power for the money. So they accepted standard configs rather than wait. Those extras are a waste in a sense, but I bet it was a deadline issue.
Not really
What that means is that the G5 X-Serves just wouldn't be shipping in time to be installed and working as a cluster by then. Apple could announce G5 1U servers next week and that would have been too late for the goals of this project.
Not really. Knowing the dealine that this system needs to be operational (not necessarily optimized) by Oct 1 for it to be included in the next super computer listing.
What that means is that the G5 X-Serves just wouldn't be shipping in time to be installed and working as a cluster by then. Apple could announce G5 1U servers next week and that would have been too late for the goals of this project.
My single G12 is as fast as...
Yes, quite right. But what´s now interesting is to see when a single processor mac will be as powerful as this cluster. In 10-15 years perhaps?
A modern chip will always be compared with the supercomputers of the past:
Now we´ve got our own modern supercomputer to compare with!
And, yes the G5´s look really good side-by-side. Funny Apple didn´t design a nice rack for them.
In a couple of years when they ditch the old to get up-to-date nodes, imagine the second hand ad...
"Previous Supercomputer seeks new homes!"...
Yes, quite right. But what´s now interesting is to see when a single processor mac will be as powerful as this cluster. In 10-15 years perhaps?
A modern chip will always be compared with the supercomputers of the past:
Now we´ve got our own modern supercomputer to compare with!
And, yes the G5´s look really good side-by-side. Funny Apple didn´t design a nice rack for them.
In a couple of years when they ditch the old to get up-to-date nodes, imagine the second hand ad...
"Previous Supercomputer seeks new homes!"...
What a COLOSSAL WASTE!!!!!!
I bet their just going to use this setup just for... for... for photoshop or something
(Just joking)
I bet their just going to use this setup just for... for... for photoshop or something
(Just joking)
Estimated Performance:
Trenchmouth:
We don't have to guesstimate Rpeak since it's just simple multiplying of known numbers! ;-)
The G5 can do 4 64bit (double Precision) FPops per Cycle due to 2 FPUs and the PPCs Multiply-Add Command, which counts as 2 FPops.
4 FPops/Hertz * 2 GigaHertz * 2 CPUs/Node * 1100 Nodes = 17.6 TFLOPS Rpeak
Now, Rpeak is just -as always- the THEORETICAL Maximum, keep that in mind! The current #2 in the Top500 "only" has 20.4 TFLOPS, too! It's just a question how close they can get to Rpeak with Rmax...
Linpack uses 64bit-Floats, but if they find Projects that can do with Single Precision (32bit) Floats, they can employ Altivec, too!
And Altivec can do 4 Multiply-Adds per Cycle! That means just using Altivec alone (think NASA, think Jet3D!) this thing could reach 35.2 TFLOPS!
Now, if they manaage to mix scalar and vector code nicecly scheduled and employ BOTH the FPUs and Altivec, we could even assume a total of 52.8 TFLOPS Rpeak (The current #1 in the Top500 has 40.9 TFLOPs Rpeak!). but i consider it unlikely they will get very close to THIS Rpeak because it's pretty tricky to keep ALL Units fed constantly without any Pipeline stalls or delays! ;-)
Now, as for the "superfluous" GFXcards: There are quite a few projects underway where Universities work with the new programmability of current GPUs (Shaders) for crunching research data. Just imagine what would happen if VUtech managed to also use THIS! ;-) These Things have AGPx8, which is 2GByte/s, a pretty handy bandwidth (though i don't know how fat the "backchannel" to the CPU/RAM/IOsubsystem is, which has been a major obstacle in the past for using GPUs for other purposes as GFXcards are traditionally designed to be "one-way"!). A GeForce FX has 51 GFLOPS (one G5 2GHz "only" has 24 GFLOPS, 8 for the FPUs and 16 for Altivec!), and the 9600 is probably just as good!
Trenchmouth:
We don't have to guesstimate Rpeak since it's just simple multiplying of known numbers! ;-)
The G5 can do 4 64bit (double Precision) FPops per Cycle due to 2 FPUs and the PPCs Multiply-Add Command, which counts as 2 FPops.
4 FPops/Hertz * 2 GigaHertz * 2 CPUs/Node * 1100 Nodes = 17.6 TFLOPS Rpeak
Now, Rpeak is just -as always- the THEORETICAL Maximum, keep that in mind! The current #2 in the Top500 "only" has 20.4 TFLOPS, too! It's just a question how close they can get to Rpeak with Rmax...
Linpack uses 64bit-Floats, but if they find Projects that can do with Single Precision (32bit) Floats, they can employ Altivec, too!
And Altivec can do 4 Multiply-Adds per Cycle! That means just using Altivec alone (think NASA, think Jet3D!) this thing could reach 35.2 TFLOPS!
Now, if they manaage to mix scalar and vector code nicecly scheduled and employ BOTH the FPUs and Altivec, we could even assume a total of 52.8 TFLOPS Rpeak (The current #1 in the Top500 has 40.9 TFLOPs Rpeak!). but i consider it unlikely they will get very close to THIS Rpeak because it's pretty tricky to keep ALL Units fed constantly without any Pipeline stalls or delays! ;-)
Now, as for the "superfluous" GFXcards: There are quite a few projects underway where Universities work with the new programmability of current GPUs (Shaders) for crunching research data. Just imagine what would happen if VUtech managed to also use THIS! ;-) These Things have AGPx8, which is 2GByte/s, a pretty handy bandwidth (though i don't know how fat the "backchannel" to the CPU/RAM/IOsubsystem is, which has been a major obstacle in the past for using GPUs for other purposes as GFXcards are traditionally designed to be "one-way"!). A GeForce FX has 51 GFLOPS (one G5 2GHz "only" has 24 GFLOPS, 8 for the FPUs and 16 for Altivec!), and the 9600 is probably just as good!
Abruptly, a week after the cluster is assembled and operational, VT requests a return:
"It has come to our attention that on a mac, you in fact cannot play counterstrike, thereby making all these things useless. Please accept our request for a full refund...."
"It has come to our attention that on a mac, you in fact cannot play counterstrike, thereby making all these things useless. Please accept our request for a full refund...."
Maybe they should hook up 1100 moniters and have a huge warcraft 3 tourney
Its just amazing I hope they get it up and running in time to be on the October list.
Its just amazing I hope they get it up and running in time to be on the October list.
Minor point to note, but they can fit 12 per rack.
You can see it best in this picture:
http://www.chaosmint.com/mac/techclusterphotos/rack3.jpg
You can see it best in this picture:
http://www.chaosmint.com/mac/techclusterphotos/rack3.jpg
Dangit, there's my missing G5!
I placed my order for a 2ghz G5 within 30 minutes of the announcement and I still don't have it!
I placed my order for a 2ghz G5 within 30 minutes of the announcement and I still don't have it!
VT is keeping all the boxes and package
The PowerMacs are standard configuration
In a year or two Apple will come out with a Power Mac with (hopefully) an IBM 980 dual core hot rod of some sort, that takes another jump in performance
Perhaps the VT students will be able to purchase these units as VT revamps (& reinvests) in their cluster
Or we could see them all on eBay
The PowerMacs are standard configuration
In a year or two Apple will come out with a Power Mac with (hopefully) an IBM 980 dual core hot rod of some sort, that takes another jump in performance
Perhaps the VT students will be able to purchase these units as VT revamps (& reinvests) in their cluster
Or we could see them all on eBay
I don't know why people are saying that use Mac users could kill their entire internet connection. That place (judging by my school) must have ATLEAST 2 0C-3's and multiple T1's as backups. I bet it has something to do with the hurricane cause we sure arn't doing it.
BEN
BEN
It's also interesting to note that the new G5 form factor allows for a more densely populated rack. That is to say, since the G5's are not rounded like the G4 towers.
Has anyone heard if these machines are stripped down on the inside?
If not, that's a HUGE, GIGANTIC, MASSIVE, ENORMOUS, TITANICALLY BIG waste of all those firewire and USB interfaces, high end video cards, even ethernet plugs! I would think Apple could've saved themselves (and Virginia Tech) a ton of cash over the span of all that wasted technology just sitting in there by leaving those elements out during assembly.
However, that would make the machines practically useless if and when the school ever decided to disassemble the cluster in the future and move the machines into other uses, such as in computer labs.
Just wondering aloud.
If not, that's a HUGE, GIGANTIC, MASSIVE, ENORMOUS, TITANICALLY BIG waste of all those firewire and USB interfaces, high end video cards, even ethernet plugs! I would think Apple could've saved themselves (and Virginia Tech) a ton of cash over the span of all that wasted technology just sitting in there by leaving those elements out during assembly.
However, that would make the machines practically useless if and when the school ever decided to disassemble the cluster in the future and move the machines into other uses, such as in computer labs.
Just wondering aloud.