So thanks for bringing the fp-precision up as an issue. I didn't think of that before. I don't know how fp is implemented in the Cell and how severe this limitation is, but I guess no one here can judge that now?
While we're at it: I think the fp-precision is a huge problem for GP-GPU as well.
MrS
There were some numbers in an recent issue of the german c't magazine.
Linpack single precision 156 GFLOP/s
Linpack double precision 9.6 GFLOPS/s
And yes, the decimal point is at the correct position!
A P4 3.4 does 25.6/7.2 GFLOP/s.
Thx for the info! If I'd be at home more often I could read the c't myself :p
So it seems that the performance of PS3 / Cell may stand and fall with the ability to use single precision as often as possible and still be accurate enough. Sounds like some serious work for the programmers!
The problem with non-IEE fp is not the decimal point, but rather the last digits. In the beginning of serious computing (1960's) there was no standard for fp and you could get quite different results from the same code, depending on which machine you'd run it. After 1 Million iterations even small differences add up (or worse: multiply).
Even with that specification IEEE 754, the standardization, though a vast improvement, does not, by itself, guarantee consistency across processors. Mostly what it does it guarantee relative levels of accuracy. For example the IBM 4381 I used for my masters project used hex based floating point. IBM had chosen this to obtain a performance improvement (which unfortunately did not fully materialize) at the cost of some accuracy.
One of the more important aspects of the impact of these things can be extrodinarily subtle. For example, you cannot (or at least I could not) represent 0.5 on that machine. What that meant was that "triggers" for program based rounding did different things depending where the numbers involved are on the number line. Ugh.
All of this leads to the need and use of the homogenous redundancy feature in BOINC where these issues can be avoided or minimized.
From your story I take it that this IBM machine had official IEEE 754 compliance?
I don't know how "hex based floating point" works, but it sounds a bit crap for now. 0.5 is one of the very few fp numbers smaller than 1 which can be represented accurately in binary, and there not even this works? :D
just a question, Does the ps3 come with an os.
If so which one. Could be a nice addition to the farm.
I do not know if the FPUs in the PS3's CPU are accurate enough to allow Einstein@home to run properly. They are optimized for fast game play and multimedia applications and are not IEEE 754 compliant. The IEEE 754 floating point mathematics standard was designed so that accuracy took precedence over performance whenever possible. This led to slow and complicated FPU designs needed for compliance with the standards. Games and multimedia applications need speed, not accuracy. They also prefer that when an overflow (when the result is too large in magnitude to represent in the chosesn floating point format) or underflow (when the result is so small that the result would round to zero even though the true result is not zero) occur that the closest value is used instead of generating an interrupt to handle it to have the program decide on how to handle the situation.
The reason that the IEEE 754 standard results in slow but accurate FPUs is that the scientific, mathematics, and accounting communities obviously will reject a standard that trades off accuracy for speed. Some scientists need as much accuracy as possible to avoid getting their work rejected or to avoid bad results. Mathematicians do not want to publish papers only to get embarassed when the math is shown to be suspect. Honest accountants need good math (unlike the ones who worked for Enron, WorldCom, or Parmalat) because they do not want their work shot down.
I stand corrected. The FPUs in the Cell Broadband Engine (CBE) (the PS3's CPU) are partially IEEE 754 compliant. They are compliant only if doubles (or 64-bit floating point numbers) are used. Single precision floats (or 32-bit floats) use the IEEE 754 data format, but do not use IEEE 754 math and do not guarantee accuracy. Einstein@home might have to be implemented using doubles to target any system using the CBE as its CPU.
If anyone is interested, I have written a guide to get the CBE simulator(it runs Linux, and has a gcc compiler) running on Gentoo(but it is generic enough for most Linux distros). It is pretty fun to play with, and might even allow the testing of the boinc/einstein client.
such things just should not be writ so please destroy this if you wish to live 'tis better in ignorance to dwell than to go screaming into the abyss worse than hell
CPU Cell will work great, as you may aware of. And PS3 is, according to CEO of SCEI, "a new platform of computer entertainment". But there is supposed to have more bigger storage than 100GB at least. The image of PS3 as a game platform comes from the two big consumers: Nintendo and Microsoft (refer here : http://pc.watch.impress.co.jp/docs/2005/0609/kaigai187.htm).
I do not know if the FPUs in the PS3's CPU are accurate enough to allow Einstein@home to run properly. They are optimized for fast game play and multimedia applications and are not IEEE 754 compliant.
{edit}: FPUs are normal. PPE in the CPU has VMX and is compatible for PowerPC 970(PowerMacG5). So will the application on PS3 be based on Mac's one? (not on intel Mac. and i cant understand why apple converted to Intel)
I hope PS3 has larger storage than at least 100GB and RAM at least 1GB. But it's hard to have those, keeping the cost same.
Anyway, because the multi-core CPU is based on 64 bit (PowerPC) architecture (and not on IEEE754?), it's sure that developers need to write application code besed to the architecture.
Somedays there will be personal computers which have Cell, but from a view of cost, perhaps PS3 will be the cheapest. The best roadmap seems to be:
PS3 release
=> code optimized for the archtecture release
=> buy PS3
=> set Linux up on it
=> crunch to BOINC
if there's a question, ask me. Being a Japanese, I can search info of PS3 for Japanese websites:)
There's also the Xbox 360 as a work cruncher to be considered. It probly already come up.
Anyway i heard that some people did convert their old xbox's to linux (V.easy) and it would work fine but slow.
From what little i know that cpu might be more related to the pc's?
Also aparently they recently uprgraded the dashboarrd (whatever that is) and it now supports loads of multitasking. Can download and play games ect.
That was on the bbc news site somewhere
The Cell architecture differs quite much from the current standard CPU designs, so while potentially being more powerful, much more effort is needed on the software side.
If you leave aside the synergistic processing elements (SPEs), you get a rather ordinary PPC-based machine. Given the projected price of the PS3, you'd be better off with a bunch of Macs than PS3s.
To make use of the SPEs, thus accesing the full power of the Cells, you don't only need a new compiler; like e.g. with processing data in the Graphics Controller you have to implement explicit, software-controlled memory accesses / tranfers. It will take some man-months of work to implement this.
Thus for Einstein@Home this would make sense for clusters of many hundreds of identical machines, exclusively running Einstein@Home, but if it's for a few dozens of users with one or two machines each, the time would be better spent by improving the Windows App by, say, 1%.
Thus for Einstein@Home this would make sense for clusters of many hundreds of identical machines, exclusively running Einstein@Home, but if it's for a few dozens of users with one or two machines each, the time would be better spent by improving the Windows App by, say, 1%.
Therefore, we should keep eyes on how many PS3 will be sold and how many people use it as a Linux-running computer. In Japan there appears no clear prediction yet around these points. But both PS1 and PS2 have been sold over one hundred million in the world, so perhaps someday PS3 also will has been sold over one hundred million.
Thanks to its price, it's possible for university students to take clusters up with PS3 if they have to calculate something which demands a supercomputer. Actually it may be easier for them to get subsidy from the university if they want PS3s, not BlueGene/L, under the condition that they won't play FinalFantasy or MetalGear...
RE: So thanks for bringing
)
There were some numbers in an recent issue of the german c't magazine.
Linpack single precision 156 GFLOP/s
Linpack double precision 9.6 GFLOPS/s
And yes, the decimal point is at the correct position!
A P4 3.4 does 25.6/7.2 GFLOP/s.
Michael
Team Linux Users Everywhere
Thx for the info! If I'd be
)
Thx for the info! If I'd be at home more often I could read the c't myself :p
So it seems that the performance of PS3 / Cell may stand and fall with the ability to use single precision as often as possible and still be accurate enough. Sounds like some serious work for the programmers!
The problem with non-IEE fp is not the decimal point, but rather the last digits. In the beginning of serious computing (1960's) there was no standard for fp and you could get quite different results from the same code, depending on which machine you'd run it. After 1 Million iterations even small differences add up (or worse: multiply).
MrS
Scanning for our furry friends since Jan 2002
Even with that specification
)
Even with that specification IEEE 754, the standardization, though a vast improvement, does not, by itself, guarantee consistency across processors. Mostly what it does it guarantee relative levels of accuracy. For example the IBM 4381 I used for my masters project used hex based floating point. IBM had chosen this to obtain a performance improvement (which unfortunately did not fully materialize) at the cost of some accuracy.
One of the more important aspects of the impact of these things can be extrodinarily subtle. For example, you cannot (or at least I could not) represent 0.5 on that machine. What that meant was that "triggers" for program based rounding did different things depending where the numbers involved are on the number line. Ugh.
All of this leads to the need and use of the homogenous redundancy feature in BOINC where these issues can be avoided or minimized.
From your story I take it
)
From your story I take it that this IBM machine had official IEEE 754 compliance?
I don't know how "hex based floating point" works, but it sounds a bit crap for now. 0.5 is one of the very few fp numbers smaller than 1 which can be represented accurately in binary, and there not even this works? :D
MrS
Scanning for our furry friends since Jan 2002
RE: RE: just a question,
)
I stand corrected. The FPUs in the Cell Broadband Engine (CBE) (the PS3's CPU) are partially IEEE 754 compliant. They are compliant only if doubles (or 64-bit floating point numbers) are used. Single precision floats (or 32-bit floats) use the IEEE 754 data format, but do not use IEEE 754 math and do not guarantee accuracy. Einstein@home might have to be implemented using doubles to target any system using the CBE as its CPU.
If anyone is interested, I
)
If anyone is interested, I have written a guide to get the CBE simulator(it runs Linux, and has a gcc compiler) running on Gentoo(but it is generic enough for most Linux distros). It is pretty fun to play with, and might even allow the testing of the boinc/einstein client.
http://lug.iit.edu/cgi-bin/wiki.pl?CBE_Simulator
such things just should not be writ so please destroy this if you wish to live 'tis better in ignorance to dwell than to go screaming into the abyss worse than hell
hello all. CPU Cell will
)
hello all.
CPU Cell will work great, as you may aware of. And PS3 is, according to CEO of SCEI, "a new platform of computer entertainment". But there is supposed to have more bigger storage than 100GB at least. The image of PS3 as a game platform comes from the two big consumers: Nintendo and Microsoft (refer here : http://pc.watch.impress.co.jp/docs/2005/0609/kaigai187.htm).
Actually PS3 doesn't differ from PC in spec (refer here : http://www.jp.playstation.com/info/release/nr_20060509_ps3.html). PS3 have RAM, HDD and even OS, Linux. But every application need to be compiled again, perhaps.
And...
Refer here: http://www-03.ibm.com/press/us/en/pressrelease/19198.wss
CPU Cell isn't just for game playing and multimadia, while PS3 may be optimized for those.
{edit}: FPUs are normal. PPE in the CPU has VMX and is compatible for PowerPC 970(PowerMacG5). So will the application on PS3 be based on Mac's one? (not on intel Mac. and i cant understand why apple converted to Intel)
{edit2}: these websites are worth to be read:
http://pc.watch.impress.co.jp/docs/2005/0208/kaigai153.htm
http://pc.watch.impress.co.jp/docs/2005/0518/kaigai180.htm
Unfortunately they're written in Japanese. Use Google or babel.
I hope PS3 has larger storage than at least 100GB and RAM at least 1GB. But it's hard to have those, keeping the cost same.
Anyway, because the multi-core CPU is based on 64 bit (PowerPC) architecture (and not on IEEE754?), it's sure that developers need to write application code besed to the architecture.
Somedays there will be personal computers which have Cell, but from a view of cost, perhaps PS3 will be the cheapest. The best roadmap seems to be:
PS3 release
=> code optimized for the archtecture release
=> buy PS3
=> set Linux up on it
=> crunch to BOINC
if there's a question, ask me. Being a Japanese, I can search info of PS3 for Japanese websites:)
There's also the Xbox 360 as
)
There's also the Xbox 360 as a work cruncher to be considered. It probly already come up.
Anyway i heard that some people did convert their old xbox's to linux (V.easy) and it would work fine but slow.
From what little i know that cpu might be more related to the pc's?
Also aparently they recently uprgraded the dashboarrd (whatever that is) and it now supports loads of multitasking. Can download and play games ect.
That was on the bbc news site somewhere
The Cell architecture differs
)
The Cell architecture differs quite much from the current standard CPU designs, so while potentially being more powerful, much more effort is needed on the software side.
If you leave aside the synergistic processing elements (SPEs), you get a rather ordinary PPC-based machine. Given the projected price of the PS3, you'd be better off with a bunch of Macs than PS3s.
To make use of the SPEs, thus accesing the full power of the Cells, you don't only need a new compiler; like e.g. with processing data in the Graphics Controller you have to implement explicit, software-controlled memory accesses / tranfers. It will take some man-months of work to implement this.
Thus for Einstein@Home this would make sense for clusters of many hundreds of identical machines, exclusively running Einstein@Home, but if it's for a few dozens of users with one or two machines each, the time would be better spent by improving the Windows App by, say, 1%.
BM
BM
RE: Thus for Einstein@Home
)
Therefore, we should keep eyes on how many PS3 will be sold and how many people use it as a Linux-running computer. In Japan there appears no clear prediction yet around these points. But both PS1 and PS2 have been sold over one hundred million in the world, so perhaps someday PS3 also will has been sold over one hundred million.
Thanks to its price, it's possible for university students to take clusters up with PS3 if they have to calculate something which demands a supercomputer. Actually it may be easier for them to get subsidy from the university if they want PS3s, not BlueGene/L, under the condition that they won't play FinalFantasy or MetalGear...