Support NeoGAF

[Flachmatuch]

Nothing is amazing if you compare it to some crazy expensive end of the line solutions or the top of the line supercomputers. Cell chip however will be sold in a device that's going to be very cheap, even compared to PCs, and is generally aimed towards mass market products. That's why it makes sense comparing it to dual G5, as they are featured in a consumer computer you can buy in any store, and are close to the top of the line in that market. For the rendering demo they showcased, Cell does hugely better job than those dual G5s, IBM had a comparision chart that displayed that, but the demo was of course suited to Cells architecture.

And this is what I'm contesting, to a degree. I don't think the Cell architecture will be so much better to general purpose computing (if even), which should be the case if Sony is comparing Cell to the G5.

 

[Vince]

And this is what I'm contesting, to a degree. I don't think the Cell architecture will be so much better to general purpose computing (if even), which should be the case if Sony is comparing Cell to the G5.

Um, ok buddy.

 

[monkeymagic]

Right now all we should be concerned with is CELL's application in a gaming machine and NOTHING touches it in that regard.

 

[MightyHedgehog]

I really wish it was May and all of this hot air would cool to something realistic and tangible instead of all of this superlative chest-pounding of specs and not-so-really video game-related demonstrations of CPU power... I suppose it can't be avoided, since anything that has to be paraded around about Cell (with the purpose of pushing the architecture itself) is going to associate itself somehow with PS3... Five months of hype to go...hopefully less if there's a February show in Japan, as was talked about by many... I want to see a range of playable games in action, already. Emotion Engine reborn indeed.

 

[gofreak]

Didn't realise intially that there are 3 pages to that article. A nice enough read. This little detail was interesting too:

When Cell runs Minor's volcano simulator, it waits for data to arrive from memory for only 1 percent of the time; the G5, in contrast, stands idle for about 40 percent of the time.

 

[fart]

noise noise noise noise noise noise noise noise noise noise noise noise noise noise noise noise noise noise noise noise noise noise noise noise noise noise noise noise noise noise noise noise noise noise noise noise TIME TO STROKE MY E-DICK!

 

[mckmas8808]

Didn't realise intially that there are 3 pages to that article. A nice enough read. This little detail was interesting too:

That is interesting. What else interesting did you read? Here and the other place I love to read what you have to say about technology, so I trust your opinon.

 

[gofreak]

Haha, I'm just learning like most people. I just thought that was a nice piece of info, because it gives us some insight into the potential benefit of the memory architecture in Cell. People often just look at the flops figures, and think that that's the best advantage Cell has to offer, but I think it only tells a small part of the story. People often wonder how Cell can be achieving 10x and 35x speedups over conventional chips, given the floating point gap is not that large, but more often than not I think the greater gains can likely be tied to the memory setup. In some ways I think just looking at floating point figures understates the potential, and the potential gains to be had.

There isn't really any other new technical detail in that article, but it's still a nice read.

 

[Flachmatuch]

Um, ok buddy.

What are you talking about? These are all vector processing tasks, this image supports just what I said. Of course the Cell is going to be good at these, but this is definitely not "general purpose"

 

[Remedy]

400 million dollars is nothing when u consider intel is opening a new 4.5 billion dollor factory to produce its new processors.

 

[Hitokage]

Mac is not a PC only if PC means a Microsoft OS, but otherwise it definitely is a "personal computer." It's not "IBM PC compatible" (well...it is if you really want it to be) but it is a PC.

When people say "PC" when referring to electronics they invariably mean an x86 box with a PC BIOS which usually implies running a Microsoft OS. When you hear "for PC or Mac" they aren't trying to be cute.

 

[Flachmatuch]

When people say "PC" when referring to electronics they invariably mean an x86 box with a PC BIOS which usually implies running a Microsoft OS. When you hear "for PC or Mac" they aren't trying to be cute.

Well a lot of them do, but a lot of them mean that thingy on the table with the tv and the typewriter that runs office and can be used to browse the internet. In the context of the article, the PC bit wasn't that incorrect imo, locking on this thing is just being pedantic, as, for the average computer user, an x86 pc and a mac (or an x86 mac hehe) is interchangeable (mostly).

 

[gofreak]

400 million dollars is nothing when u consider intel is opening a new 4.5 billion dollor factory to produce its new processors.

You're comparing fabrication investment to R&D investment?

Intel's investments in fabs is still going to be larger than the investment in fabs for Cell, but this is just one processor we're talking about.

 

[Pimpwerx]

What are you talking about? These are all vector processing tasks, this image supports just what I said. Of course the Cell is going to be good at these, but this is definitely not "general purpose"

What's the point, even? Cell isn't built for that purpose, so the point is moot, init? PEACE.

 

[Flachmatuch]

What's the point, even? Cell isn't built for that purpose, so the point is moot, init? PEACE.

Yep, I agree with that, except that Sony is trying to spin it that way by comparing it to the G5, that was all I wanted to say. Cell is great for well-known software pipeline architectures, but Sony hypes it as generally better than everything else, which is not that true.

 

[gofreak]

Yep, I agree with that, except that Sony is trying to spin it that way by comparing it to the G5, that was all I wanted to say. Cell is great for well-known software pipeline architectures, but Sony hypes it as generally better than everything else, which is not that true.

What would you like them to compare it to? They've compared it now to G5s, Pentium 4s, GPUs..I don't see what's wrong with that at all.

And this isn't Sony, by the way, but IBM.

 

[Flachmatuch]

What would you like them to compare it to? They've compared it now to G5s, Pentium 4s, GPUs..I don't see what's wrong with that at all.

And this isn't Sony, by the way, but IBM.

Well for starters they should compare it in other areas, not just easily parallelizable stuff. For example branch heavy stuff, if we're talking about games, I'd like some AI comparisons, like chess, IBM is good at that too I'm not saying it's not going to be fast, it's just that it's primarily great for games and simulations, traditional HPC things, but not for everything, and it's primarily Sony spinning it as something to replace traditional general purpose computers (just like they said with the PS2.)

 

[kaching]

It's a shame that the discussion over this article has to be buggered by the OP's choice of thread title. It's one of the more balanced, approachable reads on Cell that I've seen. It doesn't give Cell a free ride, highlights strengths and weaknesses, and does it in a way that most layman can probably understand.

 

[gofreak]

Well for starters they should compare it in other areas, not just easily parallelizable stuff. For example branch heavy stuff, if we're talking about games, I'd like some AI comparisons, like chess, IBM is good at that too I'm not saying it's not going to be fast, it's just that it's primarily great for games and simulations, traditional HPC things, but not for everything

No one is saying otherwise. But it was designed for certain things, and these are things they're going to show off, obviously.

As for AI, in particular, I think it mostly depends on what techniques you use. "AI" isn't one technique, but a plethora of them, some of which could work very well indeed on SPUs, others might have you looking toward the PPU or toward more creative solutions. I think you're considering here just the kind of simple approach common in a lot of games, a procedural approach with lots of if statements, but that's just one of many options (and perhaps an increasingly unattractive one as AI becomes more complex). That said, I do think it is an option for SPUs to handle some or all of that kind of code, but it would require a different approach to get the best performance possible.

On a general comment about branching (and relevant to the last comments above), I think you might be surprised and how much you can do to avoid branching and still get the same work done (not that SPUs can't branch - they can - they just don't have hardware branch predicition). Yes, it's less intuitive when you start boiling if statements down to algebraic expressions and the like, but branching logic or perhaps more accurately, the equivalent of it, need not be incompatible with SPUs. More often than not, you'll probably be able to do the same thing without a branch as with it.

If you're interested in some discussion of AI on Cell, including some particularly game-orientated commentary, this thread at IBM is good, and includes commentary from folks at IBM, including some of the top guys who worked on Cell:

http://www-128.ibm.com/developerwor...reeDisplayType=threadmode1&forum=739#13749192

and it's primarily Sony spinning it as something to replace traditional general purpose computers (just like they said with the PS2.)

Kutaragi has spoken about a shift in computing in the home, but talk of competing with Intel/AMD or being "better than pentium" etc. does not necessarily mean replacing desktop processors, given the presence of these companies and technologies in other markets - the markets Cell is targetting. That said, if you look at Intel's longer term roadmaps, for example, you might see things that look somewhat familiar..

 

[Flachmatuch]

No one is saying otherwise. But it was designed for certain things, and these are things they're going to show off, obviously.

As for AI, in particular, I think it mostly depends on what techniques you use. "AI" isn't one technique, but a plethora of them, some of which could work very well indeed on SPUs, others might have you looking toward the PPU or toward more creative solutions. I think you're considering here just the kind of simple approach common in a lot of games, a procedural approach with lots of if statements, but that's just one of many options (and perhaps an increasingly unattractive one as AI becomes more complex). That said, I do think it is an option for SPUs to handle some or all of that kind of code, but it would require a different approach to get the best performance possible.

On a general comment about branching (and relevant to the last comments above), I think you might be surprised and how much you can do to avoid branching and still get the same work done (not that SPUs can't branch - they can - they just don't have hardware branch predicition). Yes, it's less intuitive when you start boiling if statements down to algebraic expressions and the like, but branching logic or perhaps more accurately, the equivalent of it, need not be incompatible with SPUs. More often than not, you'll probably be able to do the same thing without a branch as with it.

If you're interested in some discussion of AI on Cell, including some particularly game-orientated commentary, this thread at IBM is good, and includes commentary from folks at IBM, including some of the top guys who worked on Cell:

http://www-128.ibm.com/developerwor...reeDisplayType=threadmode1&forum=739#13749192

Yep, tbh that is exactly what I thought, simulation type (bottom-up) AI is pretty well suited for the Cell, and lower level stuff (neural networking) even more so (but I think for that to be of any use, you'd need a lot more than a single Cell.) (I guess I'm just a bit jaded because this is related technology to what we were working on (a general purpose data extraction/processing software pipeline architecture) with my previous company, which went bust, didn't even have the money to get the patent I worked on through all the procedures.)

I think the Cell is pretty good for any kind of software pipeline (it's designed for those after all), and gaming can mostly be broken down to those kinds of problems.

Kutaragi has spoken about a shift in computing in the home, but talk of competing with Intel/AMD or being "better than pentium" etc. does not necessarily mean replacing desktop processors, given the presence of these companies and technologies in other markets - the markets Cell is targetting. That said, if you look at Intel's longer term roadmaps, for example, you might see things that look somewhat familiar..

Yep as I have mentioned in a previous post, this is a general and pretty obvious trend (and I have more trust in organic development as opposed to genius one-shot solutions.) But I seem to recall that Kutaragi did in fact say that Cell based PCs will pwn the world (after saying the same thing for PS2 based workstations.) Might not remember correctly though.