How powerful are you expecting Cell to be

[Panajev2001a]

If you're one of those who actually believe SH4 sustained 900MFlops in your average games, then you should believe Cell chip will be 80-90% effective too.
So 125GFlops would come from a BE variant with 150GFlop max

Na i would say a chp is anywhere from 40-70% effective depending on design .

Do you really think that you're average game codebase contains anything like that density of fp instructions?

Even in vertex transform code it's unlikely you can sustain even that. Lets take DC as an example. It's possible to do a simple vertex transform on DC in 12->13 CPU cycles, unfortunately the TA takes at least 26 CPU cycles to accept a simple tri. So thats 4 dot products a reciprocal and a 4 component multiply 33flops/vertex at a rate of either 13 cycles (if you don't actually write the result anywhere) or 26 if you do. Peak would be 8 flops/cycle so in this real world (although extreme) case your getting more like 1->3 flops per cycle.

So even in an extreme case your only seeing 12%->38% CPU FPU utilisation. And contarary to what some cell fanatics here would have you believe the majority of game code is nothing like that flop intensive.

I am not worried: APUs work as fast in FX/Integer math as in FP math ( according to IBM patents and Suzuoki's patent ).

 

[Guden Oden]

BUUUUUTTTT, Erp... You're talking about the majority of game code NOW!

What if, in some rosy-shimmering future, game code for Cell for, say, an RTS or such is subdivided into say, four main branches for geometry tesselation+stuff, game engine stuff+physics, player unit pathfinding/AI and enemy unit pathfinding/AI, each running on their own PE (assuming 4 in the BB). Separate APUs cound handle a unit each in round-robin fashion as sub-threads, and pathfinding and AI could be massively more complicated than it is now with the extra processing power afforded by a chip like the BB.

Remember Total Annihilation? Great RTS for its time, absolutely stinking-awful pathfinding. It had a true 3D terrain system, but it was very easy for units to get clogged up inside your base, in chokepoints on the map, or simply stuck on an incline too steep for the vehicle to get over. When the player clicked on the map, the units selected all turned and tried to head off in the new direction. Works great with planes, but poorly for land-based vehicles, especially of mixed types. They invariably got clogged all over again.

Of course, this game was new in 97, and processors were super slow back then. That's why I expect greater things of coming games. I think game code of tomorrow will be more FP intensive simply because it CAN...

 

[ERP]

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

The majority of things you wnat to see better, AI, Path finding are not FPU bound problems, they are search problems, they are constrained predominantly by random access to large central datastructures.

After graphics physcis is the other major FPU hog, but it's much harder to get the significantly more efficiency.

Yes I agree future games will use more fpu ops thn previous games, but I think that the balance will actually shift towards more AI ops relative to the total frame time.

 

[ChryZ]

"How powerful are you expecting Cell to be"

IMHO feasible ... just thinking about all the variables Sony needs to consider to make a successful mass market product, makes my head spin.

 

[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.

 

[Wunderchu]

failed project like the PS2

if PS2 was a "failed project", I'd like to see what that PS3 will be like, if it is a success

 

[Guden Oden]

And when I think about the PS3 becoming another failed project like the PS2

What an utterly bizarre person you are!

 

[thop]

What was the quote Deadmeat had in his signature again?

 

[vrecan]

he's not the only one I thought the ps2 was a horrible failure on the hardware side... luckly it has some great games.

 

[archie4oz]

I'm more curious to how you folks are quantifying 'success' or 'failure' from a 'hardware side'... :?

 

[qwerty2000]

Well I'm still looking at 1tflop for the cpu and 512gflops for the gpu.

 

[Spidermate]

And when I think about the PS3 becoming another failed project like the PS2

What an utterly bizarre person you are!

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.

Now,Sony is starting the project once more,but they are actually building it this time and with IBM at that.It serves the same purpose as the original GScube but with a Cell chip instead of multiple motherboards,which is how Sony came up with the idea for the Cell technology.With assistance from IBM and Toshiba,Sony may succeed with this project.

But,you all have to admit,for a console that is almost two years older than the Xbox,even Sony's failed project holds its own with some of its specifications in which the Xbox has not surpass or reached to this day.

Although this is very old information,here's something you all should read:
complete source
"Sony is already ramping up from 16 to 64 GS chips. And the PS3 will be made even more powerful, possibly by merging the parallel GS processors onto a superchip. (Sony has contracted the job to IBM.) But until the chip is built, the 16-processor GScube is enough to get at least some coders excited about sticking with Sony."

 

[Spidermate]

I'm more curious to how you folks are quantifying 'success' or 'failure' from a 'hardware side'... :?

Well,Sony was looking for 1000 fold and only got 300 fold (which is still good).It was still a failed project,however.But like I said before,only Toshiba and Sony were working on the project at the time,and the workstation didn't even make it into construction at the time of the PS2's development.IBM is on the project this time.They are the ones that will be building the workstation and the OS with Sony supplying the other stuff.That is why I say they might succeed this time.

 

[Vince]

The majority of things you wnat to see better, AI, Path finding are not FPU bound problems, they are search problems, they are constrained predominantly by random access to large central datastructures....

Yes I agree future games will use more fpu ops thn previous games, but I think that the balance will actually shift towards more AI ops relative to the total frame time.

So, how exactly are the - what's that phrase of yours, "Cell fanatics" - wrong when they point towards massive concurrency?

What are you going to do, keep throwing resources at singular (or near singular) scene based objects? I think games of the future are going to be about pervasive worlds filled with interactivity and gameplay potential, not the seemingly perpetual evolution of Doom and the corridor.

You're comment strikes me as totally odd, I'm sorry. And here's why so you can explain why I'm wrong... Why should we keep throwing ever greater amounts of potential logic at acclerating something which (as you stated) will never pass a rough bounding around ~40%? Why not accelerate N entities with the potentiality? It's not very likely that any of the processing elements viewed singularly will be underpowered. In my eyes, the future isn't in accelerating a single data set in a smaller timeframe, but solving a plurality of them in parallel during the timeframe.

Where am I wrong?

 

[ERP]

The majority of things you wnat to see better, AI, Path finding are not FPU bound problems, they are search problems, they are constrained predominantly by random access to large central datastructures....

Yes I agree future games will use more fpu ops thn previous games, but I think that the balance will actually shift towards more AI ops relative to the total frame time.

So, how exactly are the - what's that phrase of yours, "Cell fanatics" - wrong when they point towards massive concurrency?

What are you going to do, keep throwing resources at singular (or near singular) scene based objects? I think games of the future are going to be about pervasive worlds filled with interactivity and gameplay potential, not the seemingly perpetual evolution of Doom and the corridor.

You're comment strikes me as totally odd, I'm sorry. And here's why so you can explain why I'm wrong... Why should we keep throwing ever greater amounts of potential logic at acclerating something which (as you stated) will never pass a rough bounding around ~40%? Why not accelerate N entities with the potentiality? It's not very likely that any of the processing elements viewed singularly will be underpowered. In my eyes, the future isn't in accelerating a single data set in a smaller timeframe, but solving a plurality of them in parallel during the timeframe.

Where am I wrong?

FWIW I do believe parallelism is the future.

I just think Cell is somewhat oversold on this forum by a few people.

When I look at the patents I just don't see the revolutionary thinking that some of the people here do. It's certainly an extreme architecture, but I'm not sure that makes it a good thing.

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.

I don't honestly know the answers, and I'm fairly excited to see how the final system looks. But a lot of people around here seem to see it as some sort of panacea for all computing problems.

 

[Paul]

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.

What's not to be excited about???

If all goes according to plan with the Broadband Engine, the thing will have a theo peak performance of a Teraflops and a TOPS, with a good chunk of e-dram at a few hundred GB/S.

Am I missing anything here? I'm being absolutely serious and non sarcastic, share your vision with me and we'll come to a agreement somehow.

 

[PC-Engine]

I'm more curious to how you folks are quantifying 'success' or 'failure' from a 'hardware side'... :?

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

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

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

 

[ERP]

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.

What's not to be excited about???

If all goes according to plan with the Broadband Engine, the thing will have a theo peak performance of a Teraflops and a TOPS, with a good chunk of e-dram at a few hundred GB/S.

Am I missing anything here? I'm being absolutely serious and non sarcastic, share your vision with me and we'll come to a agreement somehow.

Write a game on a PS2 and then think about the architecture that's proposed.

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).

It's certainly possible that Sony IBM and Toshiba may well have addressed some of the limits in the PS2's architecture, but I'll wait until I seee the final box before I make that determination.

 

[Wunderchu]

I just don't see the revolutionary thinking that some of the people here do.

What the engineers lack is purity

--A year has passed since the release of PlayStation2. Is everything under full sail?

Kutaragi: Partly yes, and partly no. However, I think one needs to have a broader view when looking at PlayStation2. Keeping track of the moves of the console itself is not an issue anymore, because the box itself has no significant meaning. There is network in the first place and the box hangs onto the network. I intend to change this equation.


--I assume you are referring to "CELL," the new microprocessor chip that you will be jointly developing together with U.S. IBM Corp and Toshiba. Won't the chip be embedded into the next generation game console?

Kutaragi: Whether CELL would be built in to the game console or not is not an essential matter. Should the era of packaging continue, I guess PlayStation3 and PlayStation4 would be worth a topic to discuss, but what I would like to stress is that the concept of packaging, or box, would disappear in the broadband era. Same thing can be said of the concept of servers and clients. A band of CELL would assume the role of the existing computer system and would establish a living organism like the real cell. World's broadband will consist of an aggregation of CELL. One CELL has a capacity to have 1TFLOPS performance and an aggregation of 1,000 CELLs would have 1P (Pets) FLOPS. The capacity of 1P is an equivalent to the information processing ability of one human being. Thus creation of another world is possible if we were able to collect CELLs that equal to the capacity of 5 billion people.


-- I heard that you took the leading role in introducing the idea to develop CELL

Kutaragi: Yes, I had been imaging it in my mind from years before. It was also my idea to dub it "CELL." Although at initial stages, I had been calling it "Saibo (meaning cell in Japanese)," I christened it with an English name "CELL" in spring 2000, when I confided my thoughts to IBM Corp. I ponder that the development of CELL will bring renovation - the first in 50 years of computer history. Nothing has changed ever since ENIAC appeared until now -- where we have Itanium. To date, network-linked computers have existed as stand-alone islands. That was not much of a problem because operating systems were unevenly distributed to each island and were interchanging data among themselves.


--So you are saying that exchanging data among stand-alone computers is not enough?

Kutaragi: What would happen if things would become even more broadband and there would be no ceilings to set limitations for the bandwidth of broadband? To be sure, there are restrictions under our current wires, but shifting to fiber optics would dramatically boost the speed of communication. We are now witnessing further development of an optical switch that has a capacity to input/output data under the form of a light signal. People would start to review the current computer architecture once such networking environment of optical communication is completed. I am not denying the high processing capability of computers that establish our current networks. Microprocessors of personal computers have reached the operating frequency of 1GHz and high-powered microprocessors are embedded onto PlayStation2. Why then can't such highly capable computers interact with each other once they are connected to the Internet? The reason is neither attributed to fiber optics nor to the "Last One Mile" task of connecting high-speed lines to households. The fact that servers and personal computers have the same LSI is the greatest bottleneck that is hobbling the realization of interaction among computers. Merely connecting one personal computer to another directly by fiber optics is easy. However, if we were to connect one personal computer to ten, what would happen to the server that positions in the center of the networking? In a case where the server is also required to function as a switchboard, we must lay out legions of clusters even when we have a centralized networking topology. Furthermore, the server would collapse should we try to shape it in the form of a complete network. The idea is the same as in the case of a server break down of the e-mail service at NTT DoCoMo. Not every single person will be able to enjoy bandwidth even if fiber optics were to spread over to all households around the world.


Topology to change

--Will CELL be a resolution to solving the bottleneck you mentioned?

Kutaragi: Exactly. CELL will transform the fundamentals of the network topology. The old mechanism functioned by reading memory data into resistors and rewriting the arithmetic into memories. In short, it was just a repetition of loading and storing. Because each cash memory differs in time of access and capacity, it worked out in such hierarchical structure as primary cache, second cache, etc. On the other hand, CELL might completely transform the concept of cache as it would drastically accelerate the speed of networking. What comes into reality is that each of the astronomical number of computers around the globe could unite to form a CELL and operate by one operating system. Each CELL would be the broadband network itself. Just to give you a picture, it is like 1,000 units of computers at one company functioning as one server. In such a networking world, one would only see the overall strength of power decline when one computer drops out and vice versa. It sounds like a human society.

[source: http://ne.nikkeibp.co.jp/english/2001/30aniv/int5_1.html ]


(I first found out about that article from this thread: http://www.beyond3d.com/forum/viewtopic.php?t=8856 )

 

[Vince]

I don't honestly know the answers, and I'm fairly excited to see how the final system looks. But a lot of people around here seem to see it as some sort of panacea for all computing problems.

Well, maybe you're right. I just am of the opinion that being from B3D we're only talking of a strict subset of the larger set of computing tasks when speaking about an IC.

And, as for Cell being a "panacea" (hehe) I'd posit that when we see an Rx00 running the same set of tasks that Cell can execute at comparable speeds then, and only then, will you'll have a valid point. And then throw in the [non]local processing aspects... While I agree with you, I think it has the potential to be a very compelling system.

IMHO, it is the wildcard of the generation -- perhaps any generation this far in the history of computing since the primordial days when the current paradigm burried the NN concept within the reseach communities of the Defense establishment. It has the greatest varience in what it can rise to greatness on and, conversely, fall over.