Support NeoGAF

[xexex]

even though the actual PlayStation3 implementation of Cell might not be revealed at next weeks ISSCC, which starts on the 6th, the basic Cell Processor Element should be/will be. edit: the Cell unveiling is actually on Monday the 7th.

The Cell PE is officially slated to run at a blazing 4.6 GHz and ( i think ) produce between 256 and 294 billion floating point operations per second (Gflops)


it will be interesting to see how many transistors a Processor Element is. and how large it is, on 90 nm. that might give us a clue as to how many Sony can STUFF into a 65 nm or 45 nm chip.

hey, I'll take 8 or 16 Processor Elements on 45nm PLEASE ^__^


most are now thinking that Playstation3 CPU will probably be a 2 PE chip. some still say 4 PEs

I don't know. Panajev what are you thinking tonight man?

 

[xexex]

btw, Hannibal at ArsTechnica just commented before going to the conference

IBM's Cell at ISSCC next week: I'll be there

2/1/2005 11:19:48 PM, by Hannibal

I just wanted to drop a quick line on the front page and let the industry types who read the site know that I'll be in San Francisco on Monday and Tuesday to cover IBM's unvieling of the Cell architecture at ISSCC. If anyone wants to get in touch with me about meeting up at the conference, email me at jstokes@mailcircuit.net.

Also, if any Ars readers in the SF area are up for it, I'm sure we could do some kind of short-notice Ars meet on Monday night. If you guys want to arrange something like this, go ahead and set it up in the discussion thread attached to this post.

Finally, I should mention that I'm really looking forward to seeing what IBM has up their sleeve with this. I'm going to refrain from doing any speculating this close to the launch, but I will go ahead and hint at some of what I actually know. I once speculated somewhere (I can't find the link right now) that IBM would probably start introducing a basic core design at the top end and then migrating that design down into other niches, just like they've done with the POWER4->970 shift. I stand by this bit of speculation.

Also, regarding rumors of potential Apple uses of the Cell processor, I don't think it's at all likely (at least not anytime soon) for reasons that will become clear in Monday's session.

 

[Diffense]

Is there a compendium of what we know so far about cell anywhere?
What is enabling IBM to achieve the performance they are hinting at?
It's certainly exciting that PS3 will have all that raw processing power AND an NVidia GPU.

 

[teiresias]

What is enabling IBM to achieve the performance they are hinting at?

Hamsters. Shhhhhh, that's all I can say, sorry.

 

[Diffense]

Why didn't intel think of that! :-o
Seriously, I'll have to see what the Cell PE is capable of to believe it.
Monday will be interesting.

 

[marsomega]

These boards will be going down monday. I can really see it.

 

[Chiggs]

If the PS3 ends up having 4 PE's, then I will be extremely impressed. The theoretical performance is astounding.

 

[Izzy]

If the PS3 ends up having 4 PE's, then I will be extremely impressed. The theoretical performance is astounding.

I don't think there'll be more than one PE, the performance of the single one is amazing enough - at >250 GFLOPS it's about equal to Xenon's CPU + GPU FLOP number; and we haven't even seen the Nvidia's PS3 GPU yet.

But, as someone once said, 'Ken is insane enough to put 4 PEs in PS3.' I just wouldn't hold my breath for it.

 

[GIR]

The 'cell' which gives the chip its name doesn't refer to the hardware, but to a virtual clump of software which roams the system looking for computing resources. The patent refers to a "cell object" - program and data - and it can even roam across LANs or WANs, to find another Cell-based device.

The Register - The Cell chip - what it is, and why you should care

 

[gofreak]

At the ISSCC: 1 PE, 8 APU Cell chip,250-300GFlop

In PS3: As above. That's enough to likely make it the most powerful CPU in any of the next gen systems by quite a comfortable margin, so what'd be the point in any more, particularly when you factor in cost etc.?

 

[Pug]

The PS3 will be a 1-8 setup.

 

[xexex]

The PS3 will be a 1-8 setup.

what do you mean by 1-8 setup ?

 

[Pug]

1 PE 8 APU's.

 

[Raven.]

In PS3: As above. That's enough to likely make it the most powerful CPU in any of the next gen systems by quite a comfortable margin, so what'd be the point in any more, particularly when you factor in cost etc.?

Well, we saw the psp, it could've been weaker and with a smaller screen and it'd have been enough, but it just wouldn't be as impressive. The prestige and the PR of the Tflop

I'm hoping for a B.E.

 

[Kleegamefan]

Yeah, I am thinking SCEI will go for the Teraflop number too @_@

I don't think any other next gen console (XENON/Revolution) will be able to touch PS3s power....

 

[xexex]

dude I want at least 6 or 7 Teraflops, not just 1. remember Sony said PS3 would be over 1000 times as powerful as PS2 @__@ hehe

 

[SantaC]

Yeah, I am thinking SCEI will go for the Teraflop number too @_@

I don't think any other next gen console (XENON/Revolution) will be able to touch PS3s power....

appearently not.

 

[HokieJoe]

Hamsters. Shhhhhh, that's all I can say, sorry.

Hehe,

 

[marsomega]

So when can we expect the info leaks?

 

[doncale]

3:00 to 4:00 pm pacific time Monday. so, I'd expect info to hit the web in less than 24 hours

 

[teiresias]

How long before the full papers and transactions from the conference are available on IEEEXplore?

 

[ThirdEye]

What's the prize

 

[Andrew2]

I assume if Dr. Peter Hofstee is the one that fully present CELL,we'll be hearing some interesting thing.

 

[evil solrac v3.0]

you know the cell is a big deal when drudge is talking about it on his sunday night show.

 

[GhaleonEB]

you know the cell is a big deal when drudge is talking about it on his sunday night show.

Why the fuck would you listen to that? Ug.

 

[doncale]

you know the cell is a big deal when drudge is talking about it on his sunday night show.

get out of town :lol

 

[evil solrac v3.0]

Why the fuck would you listen to that? Ug.

he's funny in a fox news kinda way. he kept going on and on about the ethics of technology like the cell being so powerful. he kept saying [DRUDGE] you got a supercomputer the size of a pin, that's how we measure them now. the size of a pin!!! then they'll put it on cameras and such and all of a sudden our rights to privacy will be eroded" the whole thing was hillarious and he was basically talking out of his ass as usual.

 

[jimbo]

"I don't think there'll be more than one PE, the performance of the single one is amazing enough - at >250 GFLOPS it's about equal to Xenon's CPU + GPU FLOP number; and we haven't even seen the Nvidia's PS3 GPU yet."

What is the Xbox's CPU and GPU flop number? I didn't think that was revelead yet.

 

[marsomega]

"I don't think there'll be more than one PE, the performance of the single one is amazing enough - at >250 GFLOPS it's about equal to Xenon's CPU + GPU FLOP number; and we haven't even seen the Nvidia's PS3 GPU yet."

What is the Xbox's CPU and GPU flop number? I didn't think that was revelead yet.

We don't really know much about ATI's GPU so much as well as NVIDIA"s yet our buddy here seems to have some type of double standard. Sony Fan boy or insider? So far I can't think of any insider who would have an in depth look at both architectures.

 

[j^aws]

We don't really know much about ATI's GPU so much as well as NVIDIA"s yet our buddy here seems to have some type of double standard. Sony Fan boy or insider? So far I can't think of any insider who would have an in depth look at both architectures.

Well, if you take the Xenon leaked diagram, then the GPU (aka R500) has 48 ALUs.

Each ALU can do 2 scaler and 8 vector ops per clock cycle concurrently.

1 ALU = 10 ops per clock cycle.

48 ALUs = 480 ops per clock cycle.

the GPU at 500 Mhz = 480*0.5 GHz = 240 GFlops.

 

[marsomega]

Well, if you take the Xenon leaked diagram, then the GPU (aka R500) has 48 ALUs.

Each ALU can do 2 scaler and 8 vector ops per clock cycle concurrently.

1 ALU = 10 ops per clock cycle.

48 ALUs = 480 ops per clock cycle.

the GPU at 500 Mhz = 480*0.5 GHz = 240 GFlops.

j^aws, in all fairness, those leaks were how long ago again? Yes, really long time ago, 2003'ish. Even then there were reports of ATI just recently taping out the chip in the December as well as being accepted as a fake. In any case, it isn't particular anything concrete to go by either since it really didn't represent what ATI was talking about a few months ago. IIRC the only thing that stood out was the Vertex Texturing, everything else was standard PC GPU'ish. We also don't know how the R600 has evolved at this point in time, which in reality, we will be seeing an early incarnation of it called R500. Then again, with some of the talk some Sony boys have been throwing out on Cell without knowing much (excluding a few, love ya guys ) I guess they could too (of course other way around) for the ATI part. Still though, how much money would you put down on those specs ?

 

[j^aws]

j^aws, in all fairness, those leaks were how long ago again? Yes, really long time ago, 2003'ish. Even then there were reports of ATI just recently taping out the chip in the December as well as being accepted as a fake. In any case, it isn't particular anything concrete to go by either since it really didn't represent what ATI was talking about a few months ago. IIRC the only thing that stood out was the Vertex Texturing, everything else was standard PC GPU'ish. We also don't know how the R600 has evolved at this point in time, which in reality, we will be seeing an early incarnation of it. Then again, with some of the talk some Sony boys have been throwing out on Cell without knowing much (excluding a few, love ya guys ) I guess they could too (of course other way around) for the ATI part. Still though, how much money would you put down on those specs ?

Think of it this way...the CELL is a highly parallel architecture (a Stream processor)...the ATI R500 is a highly parallel architecture (also can be considered a Stream processor)...now given a similar die size and process e.g. say 90nm, would you expect these clever engineers to extract a similar amount of computing FLOPS for CELL and the R500?

CELL ~ 250-300 GFLOPS
R500 ~ 250 GFLOPS

Say both have similar die sizes and are on 90nm.

 

[Izzy]

j^aws, in all fairness, those leaks were how long ago again? Yes, really long time ago, 2003'ish. Even then there were reports of ATI just recently taping out the chip in the December as well as being accepted as a fake. In any case, it isn't particular anything concrete to go by either since it really didn't represent what ATI was talking about a few months ago. IIRC the only thing that stood out was the Vertex Texturing, everything else was standard PC GPU'ish.

Stood out? Vertex texturing is the part of SM 3.0 specs.

 

[marsomega]

Stood out? Vertex texturing is the part of SM 3.0 specs.

Whoops sorry, just got up. ( Disappointed in no CELL info blow out any where. ) Wait a minute, Vertex Texture fetching, that is part of the SM 3.0 Specs?

Think of it this way...the CELL is a highly parallel architecture (a Stream processor)...the ATI R500 is a highly parallel architecture (also can be considered a Stream processor)...now given a similar die size and process e.g. say 90nm, would you expect these clever engineers to extract a similar amount of computing FLOPS for CELL and the R500?

CELL ~ 250-300 GFLOPS
R500 ~ 250 GFLOPS

Say both have similar die sizes and are on 90nm.

It is decent "speculation", but being the ATI fan it doesnt please me as something worth my time in speculation and graphic whoring about it. Tee hee

By the way, anyone seen this yet?

Presentations:

-A Streaming Processing Unit for a CELL Processor (STI)
-The Design and Implementation of a First-Generation CELL Processor (STI)
-A Double-Precision Multiplier with Fine-Grained Clock Gating Support for a First-Generation CELL Processor (IBM)
-A 4.8GHz Fully Pipelined Embedded SRAM in the Streaming Processor of a CELL Processor (STI)
-Clocking and Circuit Design for a Parallel I/O on a First-Generation CELL Processor (Rambus)

There are some further details, I am not 100% sure if got them right, so beware:

-Operation of the 16TFLOPS rack prototype will be demo'd
-The manufacturing process of announced CELL SOI technology will be 90nm, not 65nm.
-Three of five announcements are done by STI.
-IBM starts CELL trial production for workstations in the first half of 2005.
-CELL powered home server (Sony) by the end of 2006.
-CELL powered HDTV display (Sony, Toshiba) by the end of 2006.

Sorry forgot the link. http://ps2.unisord.com/

 

[Fafalada]

j^aws, in all fairness, those leaks were how long ago again? Yes, really long time ago, 2003'ish.

Not quite 'that' long ago - it was shortly after GDC 2004 that the leaked diagram appeared on the web. And the layout in the diagram is quite clearly reaffirmed by the Xenon patent uncovered back in early December.

As far as numbers go though - publically available stuff is obviously circumstantial evidence (unless you can confirm the validity of the leak yourself ). But same applies to any "known" PS3 numbers too, hence at least the comparisons jaws made are in the same context.
Admitedly they could still be completely wrong in the end.

 

[marsomega]

Not quite 'that' long ago - it was shortly after GDC 2004 that the leaked diagram appeared on the web. And the layout in the diagram is quite clearly reaffirmed by the Xenon patent uncovered back in early December.

As far as numbers go though - publically available stuff is obviously circumstantial evidence (unless you can confirm the validity of the leak yourself ). But same applies to any "known" PS3 numbers too, hence at least the comparisons jaws made are in the same context.
Admitedly they could still be completely wrong in the end.

That was the point.

 

[Izzy]

Yeah, shader model 3.0 supports texture lookup in the vertex shader using texldl. The vertex engine contains four texture sampler stages, named D3DVERTEXTEXTURESAMPLER, D3DVERTEXTEXTURESAMPLER+1, D3DVERTEXTEXTURESAMPLER+2, and D3DVERTEXTEXTURESAMPLER+3 (these are distinct from the displacement map sampler and texture samplers in the pixel engine).

 

[marsomega]

Yeah, shader model 3.0 supports texture lookup in the vertex shader using texldl. The vertex engine contains four texture sampler stages, named D3DVERTEXTEXTURESAMPLER, D3DVERTEXTEXTURESAMPLER+1, D3DVERTEXTEXTURESAMPLER+2, and D3DVERTEXTEXTURESAMPLER+3 (these are distinct from the displacement map sampler and texture samplers in the pixel engine).

Saying something "supports" something is vague and could mean anything just like how NVIDIA supports 3DC right? ;-) There is something specific about the R500/R600 concerning this that or I'm confused. I'll look at it later.

EDIT. Think I'm still asleep. Didn't read the info right (concerning the info i just posted). Anyways, if the new info (if new) is true, we won't have any PS3 Dev Kits till the end of the year. Perhaps most likely we won't see any real time PS3 demos till then?

 

[Izzy]

Saying something "supports" something is vague and could mean anything just like how NVIDIA supports 3DC right? ;-) There is something specific about the R500/R600 concerning this that or I'm confused. I'll look at it later.

You have mistaken it for something else - check out 'the leaked' XBox 2 specs to find out what!(hint - it has something to do with the GPU and the main RAM)

 

[j^aws]

...
It is decent "speculation", but being the ATI fan it doesnt please me as something worth my time in speculation and graphic whoring about it. Tee hee
...

Well EVERYTHING is speculation until final shipping product! ...And my point being those numbers weren't completely pulled outa me A$$!

 

[marsomega]

You have consufed it with something else - check out 'the leaked' XBox 2 specs to find out what!(hint - it has something to do with the GPU and the main RAM)

Nah, it couldn't have been that. I'm so sure of it. I'll look at it later. I spent enough morning time playing on the forums, I got to get ready. Tee hee

EDIT. There should be a sticky CELL thread already. There are what, 3 or 4 current ones in rotation?