March 31 Cell thread

[Vince]

I remember when the comment of 1tflop had just come out. How the tune has changed . How the tune has changed.

How has the tune changed? IMHO it's gone from outright statements of impossibility to sheer plausibility and probability - unless your part of the vocal group who staill maintains it's impossible. Take note that many of these same folks also argued that Cell wasn't for PS3 but a "server processor". It shall indeed to interesting to see what actually occurs and compare it to what people here have said - alot of people will eat crow... either way.

 

[jvd]

Well vince some things don't change do they

 

[David_South#1]

Hi.

I'm probably one of those, to which Vince is refers. At least I fit the description, "people who believe that a ‘Broadband Engine’ is a Server oriented processor."

Anyhoo, my only defense the BE looks like an IBM Multi-Chip Model (Known as an MCM). And the real reason is that I'm just an average guy who doesn't understand the need for 4 Processor Element (each composed of up to 32 processing units, totaling 128 processing units to run a game???). One Processor Element seems ambitious enough. (I expect 1-2 PE units for Emotion Engine like purposes. And perhaps 1 ViSualizer unit.)

Note: Also wanted to remind people that the Teraflop performance mark for a "Cell" unit (whatever it might be) is and has been, by official record a year 2010 target. Not a 2005 or 2006 number.

As a token to any that may be riled by such comments, I have made one change in my expectations. I was adamant that whatever processor the PS3 arrived with, if it was in 2005, production would start with 90nm silicon. I acknowledge the group efforts make 65nm to be wholly possible.

*Just checking for news*

 

[V3]

well for them to have a vpu to make up the diffrence of a tflop from a 256 gflop cpu would be a huge departure from their current ideas.

Aren't today VPUs built from stream processors ? It doesn't matter what you call them, as long as you have resources to do the job, its not that different.

 

[V3]

Anyhoo, my only defense the BE looks like an IBM Multi-Chip Model (Known as an MCM). And the real reason is that I'm just an average guy who doesn't understand the need for 4 Processor Element (each composed of up to 32 processing units, totaling 128 processing units to run a game???). One Processor Element seems ambitious enough. (I expect 1-2 PE units for Emotion Engine like purposes. And perhaps 1 ViSualizer unit.)

Well that maybe, but the current assumption is each processor element contained 9 processors + DMA. Giving a total of 36 processors.

Like I said, if you open up today VPUs, you would see as many stream processors. We have move on quite far from the original fix pipeline or from early programmable pipeline. And this move will keep going.

 

[Deadmeat]

...

IMHO it's gone from outright statements of impossibility to sheer plausibility and probability -

It has gone the other way around, with some hardcore Sony fans like Panajev admitting that PSX3 probably won't do teraflop and lowered his expectation to 256 GFLOPS...

 

[Panajev2001a]

Deadmeat, I think a number between 256 GFLOPS and 512 GFLOPS is more accurate if you want to put words in my mouth.

 

[jvd]

Deadmeat, I think a number between 256 GFLOPS and 512 GFLOPS is more accurate if you want to put words in my mouth.

peak or sustained

I believe 256 gflops sustained in most cases

 

[Vince]

Re: ...

It has gone the other way around, with some hardcore Sony fans like Panajev admitting that PSX3 probably won't do teraflop and lowered his expectation to 256 GFLOPS...

Since he's indicative of... what exactly? I don't think he's thinking is entirely correct and many of his statements concerning, say, the BE, have bordered on almost appologetic and insanely reductionist as if he needs to find something everyone will agree with today. I just don't agree with at all. Not to mention your record wrt these things hasn't exactly been in the ballpark... Yet, if you're going to talk about the technological ability to do this, it's most certainly there and the prerequisites are as well. Lets just wait and see.

 

[Megadrive1988]

256 GFLOPs sustained might mean much higher peak FP performance. like say 512 GFLOPs ~ 1 TFLOP, since there is no way sustained performance could be as high as peak performance. It was assumed that 1 TFLOP was peak, not sustained.

25% of peak performance might be sustained. perhaps higher, or lower.

anyway. looks like we'll have to wait until June

 

[Paul]

Anyhoo, my only defense the BE looks like an IBM Multi-Chip Model (Known as an MCM).

A weak defence if you can still state this in the wake of the fact that SCE and Toshiba have entered into a joint development agreement with Rambus to impliment Redwood and XDR as the bus and memory interfaces with the Broadband Engine.

Note: Also wanted to remind people that the Teraflop performance mark for a "Cell" unit (whatever it might be) is and has been, by official record a year 2010 target. Not a 2005 or 2006 number.

No.

And the real reason is that I'm just an average guy who doesn't understand the need for 4 Processor Element (each composed of up to 32 processing units, totaling 128 processing units to run a game???).

And I have no idea why SCE is shoving billions into PS3 IC when they could just create something around 2X that of the Xbox and sell millions. Then again, Sony is always pushing barriers, look at PSP spec.

 

[David_South#1]

I may not have worded “my defense†very well.
I have a belief that “Today is tomorrows beginningâ€.
Would you bet all out on a processor technology that you haven’t tested?
My point. That whatever is coming is based on something present.
The BE engine is most like based on the MCM.
Each Element of an MCM has 4 processing cores.
The exact same number Cell would likely have in this size of silicon.
The memory architecture is also very similar.

But please feel free to explain how Rambus and XDR in some way contradict this connection.
Thank you.
(Sincerely, I would like to understand.)

“No.†And that somehow disproves what I’ve said???
Step back and work with me here. Please… OK?

"When the processor is ready around 2005, the installation will begin, and I hope that 100 million to 200 million processors will . . . form what we can really call a broadband network by 2010," Kutaragi said.
http://www.eetimes.com/story/OEG20010313S0113

Simple, yes. Not very direct, true.
But it starts here with Thee network that has been Sony’s goal for “cell†from the moment it was conceived. Sony even has an internal company called Sony Broadband Networks for which the cell is being created.

What about performance. After all we live in a real world and not just online.
What is the best FLOP performance of a chip right now?
Especially when you consider a “simple general purpose†processor like cell and not some dedicated processor like a graphics card. Think about that.

I mean just take a look at Deep Blue & BlueGene/L, the source of cell processor theory. In late 2005 BG/L will only do 110Teraflops and yet be composed of a few thousand chips. Why don’t they just sell them 110 PS3’s instead?

If IBM’s most advance projects will only be doing that then what of other parallel news like this…10GHz Multicore chip will = 1 teraflop in 2010
http://zdnet.com.com/2100-1103-5068435.html
One Teraflop in 2010? Using same design ideas that Cell follows??
Seems kind of lame that way back in 2005 they could have just bought a PS3.

Intel has a chip they call a Terahertz processor. But really it’s a program to research and produce theoretical peak they figure exists. Somewhat like the long term goals of cell.

Anyhow while I’m not a tech genius I feel people should look around and base understanding of what they know. So unless you see some real evidence that cell doesn’t have a 2010 goal, take a moment and consider what I’ve said.

 

[London Geezer]

Dont know about you guys, but everytime i read/hear the "network computing" thing, and Kutaragi saying "100-200 million chips connected together WILL BE ENOUGH" i just think "Enough for WHAT exactly? Are 100 million PS3 going to turn themselves simultaneously on in the middle of the night (or day depending on where in the world u r) and do a Skynet on us? Is PS3 the antichrist? Are we going to die just by spending £300 for a console?!".

Sorry it's early in the morning here.

 

[Panajev2001a]

I mean just take a look at Deep Blue & BlueGene/L, the source of cell processor theory. In late 2005 BG/L will only do 110Teraflops and yet be composed of a few thousand chips. Why don’t they just sell them 110 PS3’s instead?

Because for the tasks they want to use BlueGene/L for they need Double Precision FP, etc...


PlayStation 3's CELL processor FLOPS rate will be obtained using Single Precision FP and using SIMD/Vector Instructions.

While processing Scalar instructions, the APU's performance goes to 1/4th of their peak potential as only one FP/FX Unit is used ( the others are put in a low power mode to save on power consumption ).

 

[Deadmeat]

...

Because for the tasks they want to use BlueGene/L for they need Double Precision FP, etc...

Of course supercomputer users are going to stick with single precision if they could get 1000 TFLOPS for only $450,000+tax instead of a couple hundred million, because the price/performance ratio is simply unbeatable.

The fact that BlueGene is not canned and IBM is still throwing hundreds of millions into it proves that CELL is not what it's cracked up to be...

 

[London Geezer]

Re: ...

Because for the tasks they want to use BlueGene/L for they need Double Precision FP, etc...

Of course supercomputer users are going to stick with single precision if they could get 1000 TFLOPS for only $450,000+tax instead of a couple hundred million, because the price/performance ratio is simply unbeatable.

The fact that BlueGene is not canned and IBM is still throwing hundreds of millions into it proves that CELL is not what it's cracked up to be...

Funny, does anyone know what it IS cracked up to be?? Other than patents and hype, not much is known...

 

[Panajev2001a]

Of course supercomputer users are going to stick with single precision if they could get 1000 TFLOPS for only $450,000+tax instead of a couple hundred million, because the price/performance ratio is simply unbeatable.

You are simply out of your mind Deadmeat.

First super-computer users, even if they wanted to use Single Precision FP, they would have a peak of 256 GFLOPS for Scalar operation ( peak for SIMD/Vector processing would be in this scenario 1 TFLOPS ).

This would mean about $1.18 Million for your 1 PFLOPS using Single Precision Scalar FP processing.

Second, let's see: the Emotion Engine has a peak of about 6.2 GFLOPS.

Well, $179 for a PlayStation 2 and $99 for the Linux Kit and say $50 for a small 8 ports 10/100 Mbps Router/Switch.

A total of $2,274 per node ( 8 PlayStation 2 consoles with the Linux Kit and the Router/Switch ).

Each node would push then 49.6 GFLOPS.

We need 20,162 nodes then for a total price of $45,848,388 and if you wait a few months the price of the PlayStation 2 will fall to $149 bringing the price down to $41,009,508.

This is much less than "a couple hundred million dollars", does it mean that all other super-computers are doomed ? It would seem so going by what you said.

This discussion is pointless though, as I already said before, because the applications that require something like BlueGene/L need the extra precision afforded by using Double Precision FP and do most of their calculations without using SIMD operation.

 

[Deadmeat]

...

Panajev, I meant 1000 of mystical 49,800 Yen Teraflop PSX3es....

 

[Panajev2001a]

Like that saves your argument .

 

[freq]

Re: ...

Panajev, I meant 1000 of mystical 49,800 Yen Teraflop PSX3es....

No wait! Deadmeat's right!!! People are paying WAY too much for those Blue-whatever fancy IBM machines!! It's mathematically provable, using DEADMEATMATH! (tm) coming soon to a lecture hall near you...

We can see it even with today's technology! I mean, people are paying 100 million + for a BlueGene computer, when you can get 10,000 Intel Celeron 2.6 GHz processors for only 10 million!! That's one petaflop of processing power that you can buy for only 10 million dollars!

...What do you mean you have to connect them together somehow? What do you mean most of the costs are in engineering? That's CRAZY talk! That doesn't work at ALL with DEADMEATMATH! (tm), which is a proven mathematical theoremitical practice that can PROVE that Sony's next generation system will only be half as powerful as the current X-Box!