Behold, CELL specs!!!! (I think it is..)

[nAo]

]Um... does that really speed anything up? Now instead of 1 memory starved processor you have 2. One CPU is already capable of handling the load. It's the memory that's the bottleneck.

No. It's the memory that CAN be the bottleneck.
IMHO on the PS2 Sony did a good job in this department, so I expect them to act wise again on the PS3.
One can always build such a patological case where even a well balanced architecture is starwed by lack of memory bandwith.
First of all we should factor in also internal memory bandwith, and it seems a CELL processor has plenty of that.
If you add a second CELL processor that doesn't mean you have to automatically double external memory bandwith.
There are really tons of different kind of processing one can do with a custom SPUs pipeline that don't need a single external memory reference once the data stream(s) is(are) set up.
It often happens that when I found a new CELL related patent application I'm pleasantly surprised to discover that STI guys addressed a lot of the shortcomings/limitations of the Emotion Engine. Almost everything a PS2 VUs coders has dreamt of (on the hardware side..)..is there. that's really nice

If anything, stuff like HyperThreading would help you overcome memory latency.

that's so true..

Multicore would only possibly help if you multiply your cache size along with the processor count.

Going multicore on CELL do mean to add more local sram..and maybe it also means adding more SPUs L1 cache (we still don't know if SPUs have caches like a couple of patents seem to suggest)

ciao,
Marco

 

[DemoCoder]

Consider this: Lets assume the PS3 GPU is Nv5d and has 24 pipelines, each can issue 2 vec4 shader instructions per cycle, and you run at 500 Mhz. That's 96glops right there. How the fixed function parts of GPU pipelines also do work (interpolators, z-buffer, etc). Adding it all up and including vertex shaders, the GPU alone probably achieves a theoretical peak 150-200flops. So CELL only needs to reach 800gflops.

 

[Fafalada]

...So CELL only needs to reach 800gflops....
Adding it all up and including vertex shaders, the GPU alone probably achieves a theoretical peak 150-200flops

While we're at it, let's throw in another 100-200GFlops of dedicated FP capable DSP for sound (preferably a proprietary chip made by a completely different company too, say, Toshiba).
And for various media decoding, I/O and backwards compatibility, we'll have a 1-2Ghz version of EE there, should top it all off nicely.

 

[pc999]

It was 6.2Gflops (theoretical, before someone points it out) in PS2 in 2000, and when you apply Moore's law it should be 64 times increased in 2006, therefore the number < 400Gflops for PS3 is unacceptable!

So anything released by that time could have at least 400 GF, Xenon included ( if lunched by now it already would be at least 200GF ), once their are both working with IBM

 

[McFly]

You guys sound like the industry could just sit down and wait a few months till they reach a specific targeted flops rating.

A better design can always be faster than a normal "just good" design, so even if both consoles CPU's are based on the same fab technology and are released at the same time, the design of the chip can still make a noticable difference and by noticable I don't meen a few percent, I meen a few times better.

Fredi

 

[London Geezer]

It was 6.2Gflops (theoretical, before someone points it out) in PS2 in 2000, and when you apply Moore's law it should be 64 times increased in 2006, therefore the number < 400Gflops for PS3 is unacceptable!

Moore's Law had nothing to do with the increase of GFLOPS in time. Unless i'm very mistaken.

 

[one]

It was 6.2Gflops (theoretical, before someone points it out) in PS2 in 2000, and when you apply Moore's law it should be 64 times increased in 2006, therefore the number < 400Gflops for PS3 is unacceptable!

Moore's Law had nothing to do with the increase of GFLOPS in time. Unless i'm very mistaken.

In the strictest meaning, nothing. But in the loose version, has something.

http://en.wikipedia.org/wiki/Moore's_law

Formulations of Moore's law

* The most popular formulation is of the doubling of the number of transistors on integrated circuits (a rough measure of computer processing power) every 18 months. At the end of the 1970s, Moore's Law became known as the limit for the number of transistors on the most complex chips.
* It is also common to cite Moore's law to refer to the rapidly continuing advance in computing power per dollar cost.
* A similar progression has held for hard disk storage available per dollar cost—in fact, the rate of progression in disk storage over the past 10 years or so has actually been faster than for semiconductors—although, largely because of production cost issues, hard drive performance increases have lagged significantly.
* Another version claims that RAM storage capacity increases at the same rate as processing power. However, memory speeds have not increased as fast as CPU speeds in recent years, leading to a heavy reliance on caching in current computer systems.

Historical analysis of Moore's law has shown that its interpretations have qualitatively changed over the years and that it has not very accurately described developments in semiconductor technology. For example, CPU monthly shows a month-by-month display of Top Processors from Intel and AMD which gives relatively little justification for believing the law continues to operate as stated.

Thus, if your performance index is the Flops number, it should double year by year. In Intel's case, it was not Flops.

 

[London Geezer]

Thus, if your performance index is the Flops number, it should double year by year. In Intel's case, it was not Flops.

Mmm i guess. But in the end, everytime Moore's Law comes up in threads, i think, Do people still believe in something a guy said in the 70's... I mean, he whole industry has changed so much, in many different ways than were predicted, so why should we care about Dr Moore's "Law", which was just a very general observation and not a prediction anyway, formulated in the 70's?
A more "accurate" way to predict future performance is to look at recent changes between a generation anf the next, of consoles, GPUs, CPUs, Memory speed...

 

[version]

can you create faster architecture?

 

[DeanoC]

Well thats not very realistic is is?

Cell = the entire CPU in PS3 case likely to be 1 PU (Processor Unit) with 8 SPU (Synegestic (stupid name) procesing unit).

Also by PU I assume you mean pixel unit? but PU is a term used in Cell don't reuse it somewhere else. If you mean shader units maybe 64 is o.k. however if you mean pixel pipelines (ROPS in NV speak) than no way your getting 64 maybe 16 would be a better estimate.
Also if your assuming the SPU are doing vertex shading than there will be no HOS unit on the GPU, it has to be done per-vertex shader, so therfore also on SPUs.

Not sure what Transport Memory is, somewhere to remember bus timetables?

To much main RAM (256Mb) and why on earth do you think deferred rendering and pixel process on CPU? Thats just bizarre...

 

[version]

you probably to run a recursive algorithm on f@cking synergesticD)
cell patent a shit

 

[Acert93]

or, another less likely but interesting strategy. release a single PE~256 Gflop PS3 by the end of 2005 to combat Xbox2 immediately. Sony continues work on Cell with IBM. working towards releasing a multi Cell-2/PE-2 multi-Tflop PS4 (say 12-36 Tflops) by the end of this decade on smaller than 45 nm process, to combat Xbox3, instead of waiting 5-7 years to release PS4 as would otherwise be, if PS3 came out in 2006 with 1 Tflop. it still leaves a 4 year console cycle which would probably still be ok for Sony. Sony has stated that they wouldn't want a 2-3 year cycle. but 4 years wouldnt be that bad. that would also help to keep console games more upto date compared to PC games. just a far out wild thought :|

Sony has a good plan with CELL, nVidia GPU, Memory, and BR. No point in rushing at all.

On the reverse side, if they released in 2005 I see them skunked. The nVidia GPU is said not to have working silicone until the END of this year and developers still do not have development units. And CELL is pretty much nowhere to be found. A BR wont be ready this year either from what I have read. This would result in rushed games and a poor release IMO.

It is better giving Xenon a year headstart and hit a homerun with the release in 2006 than trying to react to MS. If Sony reacts to an early MS release by rushing their own release then they have already lost. None of the news releases or any of the info here indicates Sony could make a 2005 release. That would be the biggest shock of the year!

But not hitting the mythical 1 TFLOP is not a reason to rush a release. So what if it does 500 TFLOPS or 256 TFLOPs??? Does that make it any less powerful? IMO 1 TFLOP is a fanboy dream--the number means nothing more than marketing and bragging. If the Xenon comes in at 100 GFLOPs and the PS3 at 200 GFLOPs I think that is a reason to be impressed (well, I would be impressed with both considering how they would both dwarf the PC).

Also, I am sure Sony wants at least 5 years out of the PS3. They make money on the games. The PS2 had some really great selling games this fall. The point of a console is to capture those 80M customers and then to milk them! The profits are at the end of the console generation, not the beginning. This is the very reason Xenon is coming out in 2005. The Intel CPU and nVidia GPU prevent a HW shrink and from MS saving money on the most profitable time of the console. The best move for them is to move on.

 

[Acert93]

Look everyone, the PS4!!


Btw, when people start adding up the GPU, I/O, CELL, and every other aspect of the PS3 to hit 1 TFLOP I want to puke. The single press release was about CELL, not the PS3, hitting 1 TFLOP. Next thing we know we will see this:

"I have a PS3 with CELL/nVidia GPU and it hits 500 GFLOPs. I also have a TV with a CELL so that is 750 GFLOPs. I also got the Sony Network center so that pushed me all the way to 1.25 TFLOPs!! Sony rocks!"

What, can we do the same with Xenon? And since the Xenon will be networked to other Xenons we can add those? How about the PC in ones house with high end GPU--does that get added?

When Sony fans get bashed for the CELL or PS3 not hitting 1 TFLOP they have in many cases brought it on themselves. PS3 and CELL are going to be MONSTERS! But the anti-Sony and anti-PS3 fan backlash on the TFLOP stuff is going to be NASTY if people keep pushing this number around as if it is the only thing that matters.

It will be great if CELL his 1 TFLOP, but it sounds right now that the PS3 with everything in the box wont hit 1 TFLOP. Does that mean it will be a disappointment? Come on, it probably will be the most powerful HW in this next gen. Everything is relative. The goal is to SELL the CELL, not keep it in R&D until 2011!

 

[Megadrive1988]

ony has a good plan with CELL, nVidia GPU, Memory, and BR. No point in rushing at all.

On the reverse side, if they released in 2005 I see them skunked. The nVidia GPU is said not to have working silicone until the END of this year and developers still do not have development units. And CELL is pretty much nowhere to be found. A BR wont be ready this year either from what I have read. This would result in rushed games and a poor release IMO.

It is better giving Xenon a year headstart and hit a homerun with the release in 2006 than trying to react to MS. If Sony reacts to an early MS release by rushing their own release then they have already lost. None of the news releases or any of the info here indicates Sony could make a 2005 release. That would be the biggest shock of the year!

But not hitting the mythical 1 TFLOP is not a reason to rush a release. So what if it does 500 TFLOPS or 256 TFLOPs??? Does that make it any less powerful? IMO 1 TFLOP is a dream--the number means nothing more than marketing and bragging. If the Xenon comes in at 100 GFLOPs and the PS3 at 200 GFLOPs I think that is a reason to be impressed (well, I would be impressed with both considering how they would both dwarf the PC).

Also, I am sure Sony wants at least 5 years out of the PS3. They make money on the games. The PS2 had some really great selling games this fall. The point of a console is to capture those 80M customers and then to milk them! The profits are at the end of the console generation, not the beginning. This is the very reason Xenon is coming out in 2005. The Intel CPU and nVidia GPU prevent a HW shrink and from MS saving money on the most profitable time of the console. The best move for them is to move on.

I totally agree with your post. the course Sony seems to be staying on is the best.

So what if it does 500 TFLOPS or 256 TFLOPs???

hey if PS3 could do that, with the graphics power to take advantage of it, I wouldn't complain

 

[version]

my idea:

ibm create 1 processor (cell) , this can to run scalar+fpu / vector / MMX instructions ,to use about same transistors,
this cpu's max transistor count 6-8 millions
16 cells in chip about 120 millions transistors + cache + memory, less than 200 millions
this is possible on 65nm

why use slow vertex and pixel shader on GPU, when cpu is fast??
4-6 times faster
and with deffered rendering must be 1000*1000 pixel shading on cell

this architect rendering the polygons to the embended memory,after cell read it and shading the pixels, then send back to edram and CRT

PU= Pixel unit a similar than ps2 pixel unit but with clipping
this is about 500t -1 million transistors

256 PU= 256 millions transistors + 32 MB embendedmemory(256m) = 512 millions in GPU

this is possible ? and fast? and good?
my opinion this is right , cell patent is bad

 

[Megadrive1988]

my dream-PS3 that launches in 2007 on 45 nm

CPU-system:
16 Processor Elements on a 4-die MCM. that is 16 'PU' CPU cores (PPC or Power based) plus 128 S|APUs. 4.5 ~ 5.0 GHz.
4~5 TFLOPs peak performance. 1~2 TFLOPs sustained performance. about 800~900x more performance than Emotion Engine.
512 MB of on-chip eDRAM

2 GigaBytes XDR-DRAM external system memory @ 100+ GB/sec


GPU-system:
Nvidia-Sony GeForce Visualizer / GeForce Synthesizer / GeForce Realizer.

96~128 unified pipelines. optimized for pixel shading, but can perform vertex shading too. Shader Model 4+ like capability & features but under OpenGL2.0 not DirectX10. 256 MB of on-chip eDRAM @ ~2 TeraBtyes/sec

2 GigaBytes XDR-DRAM external graphics memory @ 200+ GB/sec

 

[Almasy]

Sony spent alot not only on the cpu design but also on the fabs .

This is a 300$ console even if they budget it at 500$ they still have money constraints and if they put int o large of a chip that runs to hot and draws to much power they will have alot of probems .

Well, from what we know of Xenon´s CPU, it´s supposedly in the 90nm process and doesn´t seem to have those issues (or at least we haven´t heard rumors about it).

I understand that the investment included fabs as well, but it´s still a pretty hefty investment and less money could have been spent if the chip isn´t significantly more powerfull than Xenon´s.

THe console has to a cost to power ratio and I dont' believe the 1tflop is going to be it . That is more of sony hype .

1 PU would offer 256GFLOPs, and offer sustained performance of what? 100GFLOPs or less? I understand that that´s still a very powerfull chip when put in context, but after such a seemingly big investment to come up with a chip that isn´t even half of what their stated target is, I feel I still would feel somewhat dissapointed.

 

[Inane_Dork]

No. It's the memory that CAN be the bottleneck.

I was working from a theoretical situation where that is exactly the case. Besides, basically every CPU made in the 5 years has been hampered by memory more than anything else.

There are really tons of different kind of processing one can do with a custom SPUs pipeline that don't need a single external memory reference once the data stream(s) is(are) set up.

It's true you can stop a lot of the memory problems by keeping the info streaming around the processors. But lots of programs are not amenable to that design. Excel, for instance. Oracle, for another.

 

[Inane_Dork]

Consider this: Lets assume the PS3 GPU is Nv5d and has 24 pipelines, each can issue 2 vec4 shader instructions per cycle, and you run at 500 Mhz. That's 96glops right there. How the fixed function parts of GPU pipelines also do work (interpolators, z-buffer, etc). Adding it all up and including vertex shaders, the GPU alone probably achieves a theoretical peak 150-200flops. So CELL only needs to reach 800gflops.

The real question is this: why do you need a CPU with 4x the FLOP rating as your GPU?

That means your GPU sucks or your CPU has numerous boatloads of power that are not really needed.

 

[Acert93]

So what if it does 500 TFLOPS or 256 TFLOPs???

hey if PS3 could do that, with the graphics power to take advantage of it, I wouldn't complain

Oops, I meant 500 GFLOPS or 256 GFLOPs.

Btw, your dream system WILL exist--it is called PS4! I doubt you could get that machine under $600 in 2007.

My dream system is Xenon PC A $300 PC that can play games? Oh yeah! Of course I have thousands of dollars in PC software and do not own a TV, so I am not your typical user A top of the line gaming machine that could do basic PC stuff would be a great little device, would be a cheap/powerful PC for my wife's needs, and would mock the Mac Mini hehehe I really hope MS does this, but we will have to wait to see. But I can see it now--companies upgrade to Xbox 2 over traditional PCs because they are cheaper and can do everything they need. And you can have a Halo party after work! Or ::ahem:: during work