PDA

View Full Version : The PS3's Cell


PkennethV
Apr 2, 2007, 11:31 PM
Ok...I'm not getting this. Why is the PS3's Cell so incredibly good at Folding? Does it do anything differently than "computer" processors? How did Sony, IBM, Toshiba (and some others....?) all of a sudden come up with something that beats the %@#$ out of the Core Duos and even xeons? Why aren't all the computers switching to the cell then? With this kind of power it should be blowing away the 360 (gaming wise)...

zap2
Apr 2, 2007, 11:38 PM
How did Sony, IBM, Toshiba (and some others....?) all of a sudden come up with something that beats the %@#$ out of the Core Duos and even xeons? .

No...only at certain tasks, the PS3 with Linux on it is simiarly to a PowerMac G5 1.6Ghz...which from my experiance is similar to a Core Solo Mini.


Not great, but not terrible considering the fact 4 years ago that was a top of the line Mac.

darkwing
Apr 3, 2007, 12:52 AM
To be fair the arch dependent aspects of the Cell in common Linux distributions are far from optimized yet. I think the main bottleneck is only 256 megs of system ram. I find it hard to believe that it's only performing as well as a G5 with half the clock speed, so if you have a source that'd be interesting.

GFLPraxis
Apr 3, 2007, 01:30 AM
To be fair the arch dependent aspects of the Cell in common Linux distributions are far from optimized yet. I think the main bottleneck is only 256 megs of system ram. I find it hard to believe that it's only performing as well as a G5 with half the clock speed, so if you have a source that'd be interesting.

Why is it so hard to believe?

A standard Linux build would not use the SPE's, so it would only use the main PowerPC core on the Cell, which, while it has some similarities to the G5, is extremely stripped down (as the Cell is designed to rely on the SPE's). The PowerPC core alone is rather slow (no out-of-order execution, less work per clock cycle) compared to a G5.

darkwing
Apr 3, 2007, 08:26 AM
Why is it so hard to believe?

A standard Linux build would not use the SPE's, so it would only use the main PowerPC core on the Cell, which, while it has some similarities to the G5, is extremely stripped down (as the Cell is designed to rely on the SPE's). The PowerPC core alone is rather slow (no out-of-order execution, less work per clock cycle) compared to a G5.

Sigh. I already explained out-of-order vs in-order execution to you. Yes ooo is better, but io works the same it's just dependent on static hints instead of runtime ones.

http://ieeexplore.ieee.org/iel4/6021/16082/00744331.pdf?arnumber=744331
http://www.glue.umd.edu/~ajaleel/publications/ajaleel-PhD-proposal.pdf

The latter is a very good research proposal on in-order vs out of order, and discusses many negative effects of out of order on the memory system.

While I'm not saying that out of order is bad, (because it isn't!) it certainly isn't required to have performance.

You're right of course on the SPEs. The Cell is dependent on those, and while they can run all kinds of programs Linux isn't inherently ready to do so. If someone could be clever enough to write a threading model for cell, then I think the ps3 could be a very powerful Linux machine, albeit held back with low amounts of memory. I can't see any reason the Cell can't run all your huge complex Linux programs in its SPEs and just move code chunks in and out as required. In a sense, you could treat this as a virtual cache system. That'd be an interesting project if I ever try to get my PhD. ;)

blitzkrieg79
Apr 3, 2007, 09:28 AM
No...only at certain tasks, the PS3 with Linux on it is simiarly to a PowerMac G5 1.6Ghz...which from my experiance is similar to a Core Solo Mini.


Not great, but not terrible considering the fact 4 years ago that was a top of the line Mac.

First off, the benchmarks didn't use any of the SPEs which basically the whole idea of Cell processor was lost. Second of all, IBM in 2008 will release an enhanced version of the Cell (much more improved double precision performance as well as support for 16GB of RAM) meant for general purpose computing which will go into the Roadrunner supercomputer.

http://www.cercs.gatech.edu/docs/SC06_Cell_111606.pdf

zap2
Apr 3, 2007, 09:34 AM
Second of all, IBM in 2008 will release an enhanced version of the Cell (much more improved double precision performance as well as support for 16GB of RAM) meant for general purpose computing which will go into the Roadrunner supercomputer.

http://www.cercs.gatech.edu/docs/SC06_Cell_111606.pdf

Point? This new Cell chip sound fine, but its not coming for a year, plus no computers(that you can buy) use IBM chips...Apple dropped their arse. We're talking about the PS3's cell chip...and the bench mark not using the SPEs, we'll if someone was using the PS3's chip in a computer it wouldn't either, its not optimized.


But aperently the G5 they used had a Gb more RAM then the PS3..here the link http://www.geekpatrol.ca/2006/11/playstation-3-performance/

madmax_2069
Apr 3, 2007, 01:39 PM
The Cell IMHO is not impressive to me. having that low amount of L2 cache with that many core's hurt the CPU of what it could do. it does have its stong points about it.

blitzkrieg79
Apr 3, 2007, 01:57 PM
Point? This new Cell chip sound fine, but its not coming for a year, plus no computers(that you can buy) use IBM chips...Apple dropped their arse. We're talking about the PS3's cell chip...and the bench mark not using the SPEs, we'll if someone was using the PS3's chip in a computer it wouldn't either, its not optimized.


But aperently the G5 they used had a Gb more RAM then the PS3..here the link http://www.geekpatrol.ca/2006/11/playstation-3-performance/

Point is that Cell is an architecture, it's not just one processor found in the PS3, it's a highly flexible architecture that can be adapted to specific needs. And Apple didn't drop their arse, it's more like IBM dropped Apple as their market share compared to what they could get out of consoles wasn't worth their hassle nor development of faster general purpose computing chips. In high-end processing IBM pretty much owns Intel so it's not like they don't know how to design superfast processors.

As far as the benchmark is concerned, again, Cell was designed to work as a cohesive unit with its PPE and SPEs, if the SPEs get disabled then of course it will give a subpar performance, I don't need benchmarks to tell me that. I could turn this around and refer to the original point of this discussion, look at the Fold performance of the Cell, pretty much nothing can touch it in the consumer market right now.

And sure you can get computers running on Cell, IBM sells Blade servers running on Cell.

zap2
Apr 3, 2007, 02:52 PM
Point is that Cell is an architecture, it's not just one processor found in the PS3, it's a highly flexible architecture that can be adapted to specific needs. And Apple didn't drop their arse, it's more like IBM dropped Apple as their market share compared to what they could get out of consoles wasn't worth their hassle nor development of faster general purpose computing chips. In high-end processing IBM pretty much owns Intel so it's not like they don't know how to design superfast processors.

And sure you can get computers running on Cell, IBM sells Blade servers running on Cell.

How much are these servers? Is anyone going to use this server as an everyday computer? Are you?

Yes IBM can make fast chips...can they do it in the consumer market no? Am I ever going to have server as my main computer? No...so why do I care about Cell. Apple has IBM inroad in the home market.


IBM didn't drop Apple...Apple dropped IBM, IBM was very happy when they got Apple to use G5 chips..it was victory over Intel a place were they didn't have anything. And now its gone. This new Cell chip might be good, but who is going to use it? Apple was the key into the consumer computer market. Their goal was to make a chip that so fast it burns Intel, and other companys will want some of the action. Now noone will use it, so what does it matter if it beat Intel chips? It not really worth the hassle, but if there was someone in the market selling them, companys would need to switch if they wanted to keep up with the times. Now there is no reason, so why bother?

blitzkrieg79
Apr 3, 2007, 03:15 PM
How much are these servers? Is anyone going to use this server as an everyday computer? Are you?

Yes IBM can make fast chips...can they do it in the consumer market no? Am I ever going to have server as my main computer? No...so why do I care about Cell. Apple has IBM inroad in the home market.


IBM didn't drop Apple...Apple dropped IBM, IBM was very happy when they got Apple to use G5 chips..it was victory over Intel a place were they didn't have anything. And now its gone. This new Cell chip might be good, but who is going to use it? Apple was the key into the consumer computer market. Their goal was to make a chip that so fast it burns Intel, and other companys will want some of the action. Now noone will use it, so what does it matter if it beat Intel chips? It not really worth the hassle, but if there was someone in the market selling them, companys would need to switch if they wanted to keep up with the times. Now there is no reason, so why bother?

Every company is happy when a certain transaction makes them some profit. IBM has very capable engineers, their processor manufacturing capability, on the other hand, is rather limited. It all basically came down to which would generate more money for them, a multi-billion console market (which currently they basically rule it as all the new systems use IBM developed processors) or a high demanding low profit margin computer manufacturer with a 2-3% market share at that time?

Even the G5 chips were originally made for IBMs own Blade servers, they were out in Blade servers at least 6 months before Apple started to use them.

JackAxe
Apr 3, 2007, 03:17 PM
Ok...I'm not getting this. Why is the PS3's Cell so incredibly good at Folding? Does it do anything differently than "computer" processors? How did Sony, IBM, Toshiba (and some others....?) all of a sudden come up with something that beats the %@#$ out of the Core Duos and even xeons? Why aren't all the computers switching to the cell then? With this kind of power it should be blowing away the 360 (gaming wise)...

A high-end GPU is about 4 times faster at folding than a PS3 Cell. So this beating thing really depends.

A Cell is similiar to a GPU. It's great for tasks like encoding video, or hardware rendering in Maya as an example, but it would not be a good choice for actually running Maya, even though it would do a better job than a GPU. There's a reason IBM uses them in servers, but not the workstations that access the servers. Apple also looked into the Cell, but rejected it, because it was not a good PPC replacement.

Anyway, a Cell would be a great coprocessor, just not a main desktop proc.

<]=)

zap2
Apr 3, 2007, 04:11 PM
Every company is happy when a certain transaction makes them some profit. IBM has very capable engineers, their processor manufacturing capability, on the other hand, is rather limited. It all basically came down to which would generate more money for them, a multi-billion console market (which currently they basically rule it as all the new systems use IBM developed processors) or a high demanding low profit margin computer manufacturer with a 2-3% market share at that time?

Even the G5 chips were originally made for IBMs own Blade servers, they were out in Blade servers at least 6 months before Apple started to use them.

I think your missing the point, IBM wants to make money(all companys do) so to get into the consumer CPU market, the went into bussiness with Apple, Mot(which ending up turning into freescale). There goal was to making chips which would out do intel, there more making Macs popular, selling lots and increasing $$ for all 3 companys involved.

Because IBM and Freescale could do that...make chips faster then intel, that would be work about for consumer computer...Apple was forced to look else were.


IBM would have loved to come up with a chip that was so fast, it forced Dell, HP ect into using PPC(and MS would need to port Windows to it, but this chipe would be so damn, they would need to, or else lose massive market share) but they couldn't...so now they don't have anyway to get into low end computer(the kind most people use.) So this new cell chip coming in 2008(if its as fast as stated) won't be able to get into the market. Consoles yes, high end servers maybe(if they want to use them there), but IBM won't even regain the market share they had with Apple for the consumer market.

4-5%(its not 2%, the lowest guess don't put it there) might not sound like a lot, but that would be pretty hard for a company to get. IBM won't ever regain the market the lost with Apple switching to Intel. There isn't anyone to take them, so now IBM is out of one more market.

GFLPraxis
Apr 3, 2007, 08:37 PM
Sigh. I already explained out-of-order vs in-order execution to you. Yes ooo is better, but io works the same it's just dependent on static hints instead of runtime ones.

Which is why the PS3 is so spectacular at Folding, but so-so when running an OS.

Lack of OOO is just one factor. The main cell core is just way slower than a G5 on its own. And I repeat, on its own. When running Linux, it is on its own. When running Folding, it is not, and that is why it rocks so hard at folding.

EM87
Apr 3, 2007, 09:23 PM
Hey guys do any of you know why Sony will not let Linux access the RSX chip? I find is a bit silly as if they did wouldn't it let the CPU perform better as it would not be running the GUI as well?