How powerful are you expecting Cell to be

[Spidermate]

I found something else you all might be interested in.It was back in 2000,though

Incredible details are emerging from SIGGRAPH in Tokyo, where the mega-powerful GScube was demoed to the global technology community for the first time. Our man in Japan brought us a report that detailed a showing by Square, a scene from The Matrix and a view of what could herald a revolutionary new era in entertainment across the board.
"First up, Square demoed at SIGGRAPH yesterday. Attendees saw a quick reel of a female character waking up on a lounger [couch] in a space-craft. The hair has been brute-forced [a CG technique] because so many strands were involved. Everything in the scene had been multi-passed to get the skin and eyes just right. Attendees reckoned Kuturagi-san absolutely loved this."Read that again. Now get a load of this. "Eon showed a concept preview of The Matrix. It was from the scene at the beginning of the movie, the hotel ambush. What we got was a city-top view of an animated figure running and jumping across rooftops, but you couldn't distinguish from the actual film. It was in real-time. It had an interactive camera and they were moving it around."Several other demos were shown, one of which was a flight simulator by Silicon Studio, another was a scene from the movie Antz. 150 fighting characters were seen beating it out, each with 8,000 polys eachWhat we have just outlined is the future of videogaming. Highly placed sources within the industry have been talking for some time about the convergence of movies and games, where players interact with film quality graphics indistinguishable from real life. As reported yesterday, the GScube n a CG development environment powered by the guts of 16 PlayStation 2s n is designed to take is wares to broadband networks.Welcome to the age of e-cinema.How all of this will become possible is unclear. But rest assured that you'll be playing movies in some way, shape or form in the next five years. Just ask Ken Kutaragi. "You can communicate to a new cyber-city," he said in an interview with Newsweek earlier this year. "This will be the ideal home server. Did you see the movie 'The Matrix'? Same interface. Same concept. Starting from next year, you can jack into 'The Matrix'." No-one knew just how right he was…

source

Do any one remember the bold print being mentioned at E3 this year and now taking shape as we speak?Sony has already found a way to make movie CGI in real time back in 2000 when PS2 was out.Now they have found a way to combine the two,and all of this was after the GScube 64 not 16 (as in CPU/s and GPU/s).If all goes well,they may just pull it off.

 

[Jov]

To tell you the truth,I don't know what to think about the Cell.Every time I start to doubt its performance,I end up reading a report about someone getting close to the performance in which Sony states this Cell will do.And when I think about the PS3 becoming another failed project like the PS2,then I think about the partnership Sony has with IBM combined with their previous partner,Toshiba.As far as I know,Sony might just pull this project off.I don't know how far it will distant from the Xbox 2 and NR,but I'm sure it'll look good.

Would you mind what you consider a failed project? If the PS2 is a failed project, then what can be said about the other two?

As for whether the PS3 will live up to the hype? If we get out of the PS3 as like the PS2, I think its quite a good jump already. Anything above that is a bonus.

 

[Jov]

PS2 is very good at a very small subset of problems. Games are more than graphics engines. And in my vision of nextgen a graphics engines are a much smaller part of the overall picture (i.e. they don't consume the bulk of the CPU time).

Agreed when the focus shift towards better physics and AI, thus Sony's approach seems to make it as programmerable/flexible as possible to allow the resources to be utilised (moved around) as the devs see fit.

 

[Jov]

Like I've said before, you could almost describe cell PU's as a super EE a core with 8 VU's. It just doesn't sound as exciting. Looking at them in this light you have to start asking just how good they are going to be at general computing tasks.

Are we not in an era that serialised processing is not giving the return as speed gets higher? In what scenario does general-computing tasks requires a very high Hz CPU where parallel solutions can't tackle?

(Memories of first year Eng, or Comp. Sci. principal of divide and conquer is all coming back to me )

 

[Spidermate]

To tell you the truth,I don't know what to think about the Cell.Every time I start to doubt its performance,I end up reading a report about someone getting close to the performance in which Sony states this Cell will do.And when I think about the PS3 becoming another failed project like the PS2,then I think about the partnership Sony has with IBM combined with their previous partner,Toshiba.As far as I know,Sony might just pull this project off.I don't know how far it will distant from the Xbox 2 and NR,but I'm sure it'll look good.

Would you mind what you consider a failed project? If the PS2 is a failed project, then what can be said about the other two?

As for whether the PS3 will live up to the hype? If we get out of the PS3 as like the PS2, I think its quite a good jump already. Anything above that is a bonus.

I already explained this on the second page. It is like a score you get on a test.Lets say you were aiming for an 'A' but got a 'C',instead.Although it is a passing score to teachers,it is a flunking score to you because it wasn't what you were aiming for.

Sony were aiming for 1000 fold (meaning one thousand times PS1) but got 300 fold,instead.It might have been good for most people compared to consoles on the market at that time,but it was a failed project to them.Not only was the PS2 under powered,but they didn't even get to build the workstation to power it.The thing was suppose to be as powerful as the 6800 GeForce (about a generation ahead of time).

This time is different.What Sony didn't get to do with the PS2 last time,they are doing it with the PS3 this time but much better.Meaning that if they do happen to succeed with this project,it could be as much as a generation ahead of next-generation in the very same way the PSP is to the GBA or DS.In their minds:"The hell with ATI and nVidia.This is the goal we are shooting for."

 

[Fafalada]

What I'm saying is transformation code is an absolute ideal for FPU usage and it only manages 40%efficiency.....

Well to be fair - some processors do fare better then that in transform code - but that's besides the point as we're still talking completely idealized situations yeah.

Peak would be 8 flops/cycle so in this real world (although extreme) case your getting more like 1->3 flops per cycle.

IIRC that was 7flops/cycle (no vector MADD on SH4, just Dot product), not that it matters much to this argument.

BTW wasn't the SH-4 capable of 1.4GLOPS peak?

It is. The 900MFlop I was being sarcastic about is commonly thrown around on the net as sustained performance for SH4.

 

[passerby]

In what scenario does general-computing tasks requires a very high Hz CPU where parallel solutions can't tackle?

In very general terms - many search algorithms.

Ok, I can split up the search in various ways, but I usually end up with an approximation, which is Ok sometimes. But several times the problems we try to process are either totally parallel-unfriendly or an approximation is unacceptable.

Example in gaming will be AI processing and path-finding.

 

[Jov]

In what scenario does general-computing tasks requires a very high Hz CPU where parallel solutions can't tackle?

In very general terms - many search algorithms.

Ok, I can split up the search in various ways, but I usually end up with an approximation, which is Ok sometimes. But several times the problems we try to process are either totally parallel-unfriendly or an approximation is unacceptable.

Example in gaming will be AI processing and path-finding.

AI I can understand, as for path finding I am sure this can be broken down. e.g. 4 targets, set one PF task per process unit.

I know their are situations requiring the results of one stage to feed into the next (crossed my mind after posting), thus serialised it, but even then it could be done in a pipe setup similar to current graphic pipelines.

I was more referring to are we at the stage where next-gen CPUs will be holding devs back from advancing AI. Thus, where Cell/BE could be potentially weak at? If not then what's wrong with Sony's approach?

(I've studied some AI and do appreciate the potential amount of processing resource required).

 

[passerby]

situations requiring the results of one stage to feed into the next... it could be done in a pipe setup similar to current graphic pipelines.

But as you already can appreciate, this is very hateful to implement with a problem such as AI processing. Maintanence is even more hateful. There are other problems that are similar. Search algorithms tend to be like that.

are we at the stage where next-gen CPUs will be holding devs back from advancing AI. Thus, where Cell/BE could be potentially weak at? If not then what's wrong with Sony's approach?

Well I'm not a game dev, just work for a research institute. I'm only saying that not all problems can be tackled with a parallel approach. Unless, as you suggested, you happen to have lots of them that are independent, so just launch all in parallel. I dunno about game programming, but where I work, till now I've seen 4 in-house problems that cannot be parallelized.

As a personal opinion I don't think a PE will be like a general-purpose CPU, just a "2nd-gen VPU". Only explanation for my reasoning is the cost and numbers. 32 GFLOPS per PE? And How many PEs in a PU? How many PUs on a die? And we are getting this for what? < 400USD? There's got to be a catch. And my personal reasoning is that it won't be like a general-purpose CPU, but a "2nd-gen VPU". I don't think devs will be running AI codes on them, among various other routines.

My opinion of course.

 

[Jov]

But as you already can appreciate, this is very hateful to implement with a problem such as AI processing. Maintanence is even more hateful. There are other problems that are similar. Search algorithms tend to be like that.

Yep, search algorithms and AI tends to go hand in hand. In fact after the first few lectures in 2nd (could have been 3rd) year AI, my initial impression of the entire field was simply search algorithms and logic (thus more decision making).

At that time the lecturer kept telling us AI was still very much at its infancy. I am not sure how much advancement 8/9 years on, but my belief back then was for true AI to happen, a machines much be able to do things (produce unpredictable results or results within a range that was too large to store possibly stored. e.g. human decisiong making), or think so to speak without the author/programmer requiring either to program it or input all the possibilities. For this to happen, my thoughts at the time was of self modifying/evolving code for this to be possible. Thus to allow the machine to learn and evolve.

The topic of learning itself is a specialised field within AI.

O'cause there is a lot more to AI than what I've just recalled.

 

[zidane1strife]

How powerful

I hope they use 45nm with all the advanced material and manufacturing technique bell & whistles, and that this can allow them to escape most heat/power performance ratio problems and provide massive & fast chips. But alas this might not be, in the event they use 65nm I hope they manage to apply the most innovative and effective techniques in manufacturing they can.

PS It'd be funny if one of the console manufactures went with 90nm(especially conservative) and the cell turned out to be 45nm(aggresive design bigger/faster than usual-capacitor less dram, other techs-, they could get lucky)... process wise it'd be as distant to the next-gen console as that console's from the xbox(.15micron)....

edited

 

[archie4oz]

Well,it's true.Sony was reaching for something that ended as a dead end because they didn't have all the resources back then.But after the PS2 launched,they found a way to make it more powerful,but it was to late.That was the GScube;the box that was originally suppose to be for the PS2.From what we know,the thing was able to do work from some CGI movies or,in short, nVidia's GF 6800.

I don't know about using the GScube to make the PS2 more powerful, it was simply an explorative research project... And I don't know where you make the connection with the GeForce 6800... The ideas behind the GScube had more in common with UNC PixelPlanes than anything Nvidia's conjured up...

Well,Sony was looking for 1000 fold and only got 300 fold (which is still good)

The 1000 fold was simply pie-in-the-sky roadmap stuff for public consumption. I certainly hope your not basing your theory of Sony's engineering goals on something as trivial as a silly PowerPoint slide...

Whether or not it's trivial to do a certain operation like mipmapping?

In silicon?

Anyway I predict the CELL cpu will be able to do 1TFLOPS peak using water cooling.

Well considering the DC even had heat-pipes I don't consider that much of an issue...

Do any one remember the bold print being mentioned at E3 this year and now taking shape as we speak?Sony has already found a way to make movie CGI in real time back in 2000 when PS2 was out.Now they have found a way to combine the two,and all of this was after the GScube 64 not 16 (as in CPU/s and GPU/s).If all goes well,they may just pull it off.

Yes, 'cause I was there (for a good reason )... And it *was* GS16 that ran all the demos... There was only one GS64 built (prototype) and it wasn't used for anything...

This time is different.What Sony didn't get to do with the PS2 last time,they are doing it with the PS3 this time but much better.

And do you mind explaining what this nebulous thing that Sony is getting to do with the PS3 that they didn't with the PS2?

 

[Megadrive1988]

basicly, I expect PS3 to be stronger than the GSCube 64 in most respects even concidering GSCube's massively staggaring amout of memory

 

[MfA]

AI might not need FLOPS, but at the same time it is poorly parallelized at the fine grained level (ie. bad code for serial processors too) but easily parallelized at the coarse grained level (each entities AI can done independently in a timestep system).

All interesting problems in simulation can be efficiently parallelized, because when we get right down to it nature is parallel.

 

[Spidermate]

I don't know about using the GScube to make the PS2 more powerful, it was simply an explorative research project... And I don't know where you make the connection with the GeForce 6800... The ideas behind the GScube had more in common with UNC PixelPlanes than anything Nvidia's conjured up...

Yes,it was,but Sony thought about the project at the time of the PS2's construction.The project just never unfolded at the time.Later,Sony found a way to combine both GS and EE to where it made a more powerful system.It doesn't really matter how accurate it was to the 6800 GF.That was an example.The point is,it didn't get pass it's prestage to power the system.The GScube came later as something for movie studios and not games.Ofcourse it's not going to work on a console.The workstation they thought about came before the GScube.The GScube was half the results.

The 1000 fold was simply pie-in-the-sky roadmap stuff for public consumption. I certainly hope your not basing your theory of Sony's engineering goals on something as trivial as a silly PowerPoint slide...

And I sure hope you aren't basing yours on some slide neither.If they thought their project failed to reach their expectations,that normally says something.

Yes, 'cause I was there (for a good reason )... And it *was* GS16 that ran all the demos... There was only one GS64 built (prototype) and it wasn't used for anything...

HUH? I was talking about the new workstation as in today.What are you talking about? I believe,though,that it was the GS64 that ran Final Fantasy:The Spirits Within demo in real time.

And do you mind explaining what this nebulous thing that Sony is getting to do with the PS3 that they didn't with the PS2?

The workstation for one and the GS/EE as the other. Very simple.

 

[archie4oz]

Sony thought about the project at the time of the PS2's construction.

Sorry, but the GScube was exploratory research post PS2 design... Much like the I-32... Partially to explore ideas proposed during the design, and partially to explore manufacturing prowess.

And I sure hope you aren't basing yours on some slide neither.If they thought their project failed to reach their expectations,that normally says something.

I base mine on conversations I've had with people who actually designed it... Just out of curiosity do you also consider Xbox and the GCN engineering failures too?

I believe,though,that it was the GS64 that ran Final Fantasy:The Spirits Within demo in real time.

Considering I worked on that demo, I can pretty much tell you that it was *not* a GS64...

The workstation for one and the GS/EE as the other. Very simple.

OK, now you're losing me... But a DTL *is* technically a workstation so you can't really say they failed or weren't able to follow through in that regard...

 

[ERP]

AI might not need FLOPS, but at the same time it is poorly parallelized at the fine grained level (ie. bad code for serial processors too) but easily parallelized at the coarse grained level (each entities AI can done independently in a timestep system).

All interesting problems in simulation can be efficiently parallelized, because when we get right down to it nature is parallel.

True to a point, but the issue with Cell as I see it for parlellising these tasks is the realtively small memory window for each of the PE's. Your not going to be able to run something that requires random access to a large (multimegabyte datastructure) efficiently on a PE.

You might be able to rearrange your datastructures to accomodate the local memory restrictions, but as worlds get bigger, and the structures more complex, it's going to get harder and harder.

 

[Spidermate]

Sorry, but the GScube was exploratory research post PS2 design... Much like the I-32... Partially to explore ideas proposed during the design, and partially to explore manufacturing prowess.

I don't think you quite understood me.Sony thought of the idea for a workstation assisting the PS2,not the GScube itself.That came after this idea.That was only partially what they originally wanted to do.

I base mine on conversations I've had with people who actually designed it... Just out of curiosity do you also consider Xbox and the GCN engineering failures too?

I can't really tell you anything about Nintendo since their 64 disappointed pretty badly.However,Microsoft is a different story.Althought they had a set goal in mind (which they did fail on),it wasn't no where near the goal Sony was shooting for.Their goals are normally revolutionary.That is a much bigger difference than those that are evolutionary.For a much more broad example,read Moore's law and compare it to Sony.Sony's project failed,plain and simple.

Considering I worked on that demo, I can pretty much tell you that it was *not* a GS64...

Someone famous in here,I see.I guess that makes the 64 version even more powerful than I thought.

OK, now you're losing me... But a DTL *is* technically a workstation so you can't really say they failed or weren't able to follow through in that regard...

Ok,lets put it this way: instead of Sony scrounging for hardware to move closer to photo realism,they go straight to the source by creating a bridge:a new version of the GScube.That way they are feeding this thing (which is the workstation) with information coming in from Hollywood movies/hardware and running it straight to the video game hardware.The information is simplified to where it can be understood on both sides.This is the first time this has ever been possible.

The EE and GS I was referring to was Sony combining them to make them more powerful than if they were working by themselves.It is how they got the GScube to do footage in real time from movies.The GScube's interior boards are nothing more than a Cell chip shrunken down to size but able to hold a helluva lot more than a few huge boards because of its APUs/cores.This chip,by itself,makes up the GScube(s),theoretically.

 

[Lazy8s]

A design ideology does not have to choose between the extremes of parallelism and serialism as if there's no middle ground. The development environment for parallel processing is currently less mature, so relying on some approaches that are more serial in nature to avoid certain present inefficiencies may lead to better effective performance. The successful design will be aware of this balance.

 

[nAo]

True to a point, but the issue with Cell as I see it for parlellising these tasks is the realtively small memory window for each of the PE's. Your not going to be able to run something that requires random access to a large (multimegabyte datastructure) efficiently on a PE.

You really don't want to access a large data structure on any console.
As long as a PU has a L2 cache and an APU can transfer data from there there is not much to worry about.