First PS3 "Supercomputer" ...

Status
Not open for further replies.
J

Jesus2006

Regular
Here they made a cluster of 8 PS3s running Linux:

http://www.heise.de/newsticker/result.xhtml?url=/newsticker/meldung/86534&words=Playstation (in german)

  • Author: Dr. Frank Mueller of North Carolina State University (NCSU)
  • overall cost below 5000$
  • one drawback: only 256 RAM (+ 256 VRAM) but they think they can enhance this - haven't opened up the PS3 yet
  • theoretically >200GFlops per CELL but only 6 SPEs under Linux usable, overall max. 1,2 TFlops (still not enough to enter the top 500 list though, but entirely scalable)
Click to expand...

Seems this time it's much easier to get a "supercomputer" out of some standard playstations than the last time :)
 
  • Like
Reactions: one
Capeta said:
Last time I checked each CELL was only capable of around 27 GFLOPS DP.:oops:
Click to expand...

Yeah, this figure is SP of course.

Still 27GFLops DP is far from bad - still better than most modern desktop CPUs.

-

A Core 2 Quad does about 21.x GFlops in Linpack (think it's SP here though?)
 
Capeta, 'supercomputing' can entail more than simply DP-necessitated operations; take for instance the current industries at which the QS20 is targeted.
 
Carl B said:
Capeta, 'supercomputing' can entail more than simply DP-necessitated operations; take for instance the current industries at which the QS20 is targeted.
Click to expand...

Of course, but if you're talking about entering the Top500 DP is everything and SP is nothing.

overall max. 1,2 TFlops (still not enough to enter the top 500 list though, but entirely scalable)

That's 1.2TFLOPS of pure BS SP...completely utterly useless when talking about Top500.
 
I think the mention of the Top 500 was really only in there for article 'flavor.' IMO even in the SP-constrained environment it operates in, the cluster can qualify as an HPC (or supercomputing) environment. You're right though that the Top 500 is not really in the cards for for the current revision.

This is a shortcoming that was acknowledged in the original article:

“Right now one limitation is the 512 megabyte RAM memory constraint, but it might be possible to retrofit more RAM. We just haven’t cracked the case and explored that option yet.” Another problem lies in limited speed for double-precision calculations required by scientific applications, but announcements for the next-generation Cell processor address this issue.
Click to expand...

http://www.physorg.com/news92674403.html

How that physorg article didn't make it's way into this thread vs that German article is honestly a little surprising. Anyway, I don't think the context of the article on the whole PS3 cluster thing (on the scale at NC State) was supposed to be in the 'impressive' category so much as just in the 'cool' category. Certainly people shouldn't think that this is more than it is.
 
Capeta said:
Last time I checked each CELL was only capable of around 27 GFLOPS DP.:oops:
Click to expand...

I said something close to that in a Wii thread i made. Strangely, I did a rough calculation of the PS3's real full maximum vector performance (don't ask me how) and i worked it out to be around (+/-) 30 GFLOPS which is good but not exactly a 21st century supercomputer.

I also said it s max power was equivalent to a 3Ghz G5.

Looking at it from a non gaming point of view....8 PS3s..WOW great, "if" it saves you money from buying Cell computers go for it.
 
Capeta said:
Of course, but if you're talking about entering the Top500 DP is everything and SP is nothing.

overall max. 1,2 TFlops (still not enough to enter the top 500 list though, but entirely scalable)

That's 1.2TFLOPS of pure BS SP...completely utterly useless when talking about Top500.
Click to expand...

Not true.

In Cell Symposium 2006, the research community also found a way to achieve DP-level result using a mix of SP and DP methods. The paper(s) suggested that this gives them up to twice the performance (of pure DP math). In effect, your old supercomputer can now run twice as fast (so to speak). They found this method because they were experimenting with Cell's SP performance.

There was also speculation on whether to modify the Top 500 supercomputer list (by adding 1 more column to reflect this technique) but I don't think it will happen.

The point is SP math can be useful in scientific computing too. The national labs won't spend money on useless machines because they have budget constraints like everyone else.

ATI-liens said:
I said something close to that in a Wii thread i made. Strangely, I did a rough calculation of the PS3's real full maximum vector performance (don't ask me how) and i worked it out to be around (+/-) 30 GFLOPS which is good but not exactly a 21st century supercomputer.

I also said it s max power was equivalent to a 3Ghz G5.

Looking at it from a non gaming point of view....8 PS3s..WOW great, "if" it saves you money from buying Cell computers go for it.
Click to expand...

Why argue about the term "supercomputers" when it's a moving target every year...

For certain problems and if programmed properly, Cell can be pretty efficient and fast. The BFS example demonstrated this rather well vis-a-vis a Woodcrest or a bona fide 128/256 CPU Blue Gene.
 
Last edited by a moderator:
patsu said:
Not true.

In Cell Symposium 2006, the research community also found a way to achieve DP-level result using a mix of SP and DP methods. The paper(s) suggested that this gives them up to twice the performance (of pure DP math). In effect, your old supercomputer can now run twice as fast (so to speak). They found this method because they were experimenting with Cell's SP performance.

There was also speculation on whether to modify the Top 500 supercomputer list (by adding 1 more column to reflect this technique) but I don't think it will happen.

The point is SP math can be useful in scientific computing too. The national labs won't spend money on useless machines because they have budget constraints like everyone else..
Click to expand...

Ok good point, however my point was that the 1.2TFLOPS figure is clearly not referring to the SP to DP conversion floating point performance. It is clearly referring to SP performance. Even if that SP>DP "hack" gets accepted by the Top500, it'll still be "only" around 400 GFLOPS of DP..nowhere near 1.2TFLOPS.
 
Capeta said:
Ok good point, however my point was that the 1.2TFLOPS figure is clearly not referring to the SP to DP conversion floating point performance. It is clearly referring to SP performance. Even if that SP>DP "hack" gets accepted by the Top500, it'll still be "only" around 400 GFLOPS of DP..nowhere near 1.2TFLOPS.
Click to expand...

But of course... The original article mentions this too... that to make it into top 500, they have to hit 2.7 TFLOPS and in double precision. So the quoted 1.2TFLOP should be understood as single precision. I don't think they are trying to mislead anyone.
 
patsu said:
But of course... The original article mentions this too... that to make it into top 500, they have to hit 2.7 TFLOPS and in double precision. So the quoted 1.2TFLOP should be understood as single precision. I don't think they are trying to mislead anyone.
Click to expand...

I don't read German so if you have the original article in English, a link would be appreciated.

I found this but there is no mention of doubling performance to 2.7TFLOPS...

http://www.engr.ncsu.edu/news/news_articles/ps3.html
 
patsu said:
Use google translate and you'll see... ;)

The article you linked to does not even talked about TFLOP.
Click to expand...

I think you're trying too hard.;)

NCSU is a US university, I don't see why I should translate a German article to know what Dr. Frank Mueller is doing. He doesn't claim anything like that glorified German piece.
 
Who's trying too hard ?

The OP linked to a German article which mentioned the 1.2 TFLOP figure as well as the 2.7TFLOP double precision cut-off for top 500 supercomputers. The information is accurate and has no BS.

If the cluster would scale one hundred-per cent, a theoretical total output of 1,2 TeraFlop/s would come out. To the introduction into the Top500-Liste of the supercomputers is not enough still for a long time. At place 500 at present a Xeon cluster with a maximum floating decimal point achievement (Rmax) of approximately 2.7 TeraFlop/s ranks - and in double accuracy. Who wants to develop its own PS3-Cluster: The sales of the Playstation 3 is in Europe to 23. March start. In Japan the console is since 11 November 2006 in the trade.
Click to expand...

The English article you linked to doesn't even have any TFLOP rating to begin with. It also showed that the German article was not mis-handled (Hey, they even value add by throwing in the theoretical TFLOPS).

Where is the BS ?
 
patsu said:
Who's trying too hard ?

The OP linked to a German article which mentioned the 1.2 TFLOP figure as well as the 2.7TFLOP double precision cut-off for top 500 supercomputers. The information is accurate and has no BS.

The English article you linked to doesn't even have any TFLOP rating to begin with. It also showed that the German article was not mis-handled (Hey, they even value add by throwing in the theoretical TFLOPS).

Where is the BS ?
Click to expand...

The BS starts where it mentions 1.2TFLOPS of SINGLE PRECISION. Scaling it 100% means 2.4TFLOPS of SINGLE PRECISON. Why are they comparing 2.4TFLOPS SINGLE PRECISION not being enough to match 2.7TFLOPS of DOUBLE PRECISION if not to sensationalize??? They would need to scale it A LOT more than 100% if they were intending to compare DP numbers.

Please explain to me where the 1.2TFLOPS comes from and how 8 PS3 = 1.2TFLOPS and how scaling 100% gets you 1.2TFLOPS.

The article I linked to doesn't mention TFLOPS of any kind because *gasp* it's from the universities own news and Dr Frank Mueller isn't making any glorious sensationlist claims of TFLOPS.
 
Last edited by a moderator:
Capeta said:
The BS starts where it mentions 1.2TFLOPS of SINGLE PRECISION. Scaling it 100% means 2.4TFLOPS of SINGLE PRECISON. Why are they comparing 2.4TFLOPS SINGLE PRECISION not being enough to match 2.7TFLOPS of DOUBLE PRECISION if not to sensationalize???
Click to expand...
I think you're kind of preaching to the choir here.

People are aware of the PS3's limitations. Don't get so upset that some guy's clustering them and throwing around th word "supercomputer". No need to get bad flashbacks from PS2 launch kay? :cool:

I tend to lookat this as cool trivia that somebody's doing it. Doesn't concern me I'll just play my games kkthx. Maybe you should just do the same?

Peace.
 
Rainbow Man said:
I think you're kind of preaching to the choir here.

People are aware of the PS3's limitations. Don't get so upset that some guy's clustering them and throwing around th word "supercomputer". No need to get bad flashbacks from PS2 launch kay? :cool:

I tend to lookat this as cool trivia that somebody's doing it. Doesn't concern me I'll just play my games kkthx. Maybe you should just do the same?

Peace.
Click to expand...

I know people are aware and I'm not upset that Dr. Mueller is talking about "supercomputer" because even a cluster of 486s can be a supercomputer.:LOL:

I'm just trying to get Patsu to explain to me what kind of math they're using here to get these TFLOPS numbers then using them to compare to actual TFLOPS from a Top500 supercomputer. So far the math from that sensational article doesn't add up. If you are going to defend an article then go ahead just make sure the math supports your defense.
 
Capeta said:
I know people are aware and I'm not upset that Dr. Mueller is talking about "supercomputer" because even a cluster of 486s can be a supercomputer.:LOL:

I'm just trying to get Patsu to explain to me what kind of math they're using here to get these TFLOPS numbers then using them to compare to actual TFLOPS from a Top500 supercomputer. So far the math from that sensational article doesn't add up. If you are going to defend an article then go ahead just make sure the math supports your defense.
Click to expand...

I have zero idea where you get your 2.4 TFLOP number from.

Did you read the translated article ? They laid out how they arrived at the numbers almost step by step.

From an external source (like here: http://www.lbl.gov/Science-Articles/Archive/sabl/2006/Jul/CellProcessorPotential.pdf), they assumed 200 GFLOP @ 3.2Ghz for 8 SPEs.

Since only 6 SPEs are active in PS3 Linux, they estimated linearly that they can get 150 GFLOPs.

With 8 PS3s, assuming 100% scaling, they claimed 8 * 150 GFLOP = 1.2 TFLOP.

It's easy to find where they make assumptions and disagree/agree.

Now that I read more closely, they even noted that Cell is "suitable only conditionally for scientific computations, which take place as a rule with double accuracy".

There is some 80 GFLOP number thrown around but the translation broke down here, perhaps refering to this (http://sc06.supercomputing.org/schedule/pdf/pap225.pdf).
 
Last edited by a moderator:
Status
Not open for further replies.

Similar threads

P
Look at this Google-cached (pulled down) PlayStation 3 page
10 11 12
Replies
224
Views
30K
Guden Oden
G
Back
Top