Cell & XBOX 360 Architecture

[poopyman67]

As soon as I came across this article, i knew something was up, so I read about the cell architecture on this site, and i thought... and i thought some more, and it just doesn't add up to me..

How does the 360 do >1tf/s w/6 cores made up of dumb downed g5s(well not necessarily dumb downed, i guess u could say "optimized" for its purpose, anways, i diverge from my point...), while a 16 core server can only(read "OMG!") 57Gf/s?

Then... they have the audacity to say that the cell is capable of 2TF/S?? What?? I read the article quite thouroughly, and with all the benefits of in-order architecture, and yada yada, I still dont see how its 20x more powerful than a 16 way server....

and @ >1/25 the cost... It seems to go against all logic and rationality....

or maybe im just plain ignorant....

the only logical explanation is that this number also includes the gpu.

thoughts?

 

[poopyman67]

just read this in the cpu forums" Hmm, Sony up to the same old BS as usual. As to my knowledge, the xbox 360 has a four fold advantage over the PS3 in regard to raw processing power. The Sony unit (CELL) has a tested output of somewhere in the area of 200-250Gflop. Now architectural differences aside (We'll get to that later), the tri-coreIBM PowerPC processor in the 360 can perform well over a Tflop (emphasis on "well over"). Some claim it to run closer in the range of about 1.2Tflop (Unconfirmed). I saw the whole presentation for the PS3, and I felt that deja vu all over again. Sony claims that it's graphics core adds on an additional 1.8Tflop of processing power for a total of 2Tflop. ??? ... ??? ... ??? . WOW! And I thought with the release of the PSP that Sony was finally ready to step up to the big boys' table... I was wrong. The absurdity of their claims can't even be measured, as a graphics core has nothing to do with the systems CPU. To put it simply, SCEI is lying, again. And simply is trying to their way out of another technology spoof (and miserably failing). Cell is no where close to the power of microsoft's processing solution, and just as last time when they claimed that their "emotional engine" was on top, it ended up about 1/10 as fast as the marketed number. This isn't to say that the PS3 is inferior to the XBOX 360, only that we have seen this before and obviously microsoft came out on top. As far as developing goes and coding, this system looks even more rediculous than the emotion CPU. Many argue that the PS2 was more powerful than the xbox when people learned how to use it, but to be honest no one (except maybe bungie) ever dug that deep into the XBOX's power. My guess from the lineup of 3rd parties this time around (a vast majority of Sony's staples hoping on the XBOX360 bandwagon) that sony has finally run out of marketing tactics to appease the consumers, and with development costs skyrocketing, I think developers will put more consideration into what system they decide to base their games. But this is really just fuzz right now. Most consumers are terribly educated and will usually just listen to whatever the cool crowd says. Last gen, Sony had overwhelming support from pop culture, which made playstation into the cliche that took out nintendo. I think already with microsoft's MTV marketing campaign and sleek design (not to mention billions of dollars) that they will push sony out of the market. But this all remains to be seen. One thing is for damn sure. CELL is no where near as powerful as the 360CPU. It's almost unethical to compare them. "-SoothingRelease

sounds pretty sensible 2 me

 

[Velk]

Originally posted by: poopyman67
the only logical explanation is that this number also includes the gpu.

thoughts?

It does, and the number is also intentionally misleading, in that the numbers are
a) theoretical
b) more than best case (100% utilization of all elements is clearly impossible, as they rely on each other for input, they even admit that themselves)
c) using non-standard test parameters.

It's the functional of equivalent of suggesting that your new video card does 180,000 3dmarks, while neglecting to mention that the numbers are based on running 320x200 resolution in 16 colors using Bob's 3dmarrk(tm), a test suite which doesn't actually have anything in common with the other testing product of a similar name.

I recommend reading the specs as "Achieves over 2.0 SonyMarks".

 

[0roo0roo]

yea sony has a track record of outrageous claims. remember the emotion engine.

 

[Azsen]

Originally posted by: Velk
It's the functional of equivalent of suggesting that your new video card does 180,000 3dmarks, while neglecting to mention that the numbers are based on running 320x200 resolution in 16 colors using Bob's 3dmarrk(tm), a test suite which doesn't actually have anything in common with the other testing product of a similar name.

LOL :laugh:

 

[RaynorWolfcastle]

OK, so I've figured out how they roughly came up with the 225 GFlop number. So the Cell in the PS3 is one PPE (which is essentially some derivative of the PowerPC core) and 7 SPEs. Now, hypothetically, if you somehow found a way to keep the execution units fed at all times and SPEs were exclusively doing SIMD multiply-adds you'd have the following scenario:
- 7 SPEs
- Each SPE is dual issue and is executing a FMADD
- Each FMADD is 3 mults and 2 adds per cycle
- Core clock of 3.2 GHz

So multiply this out and you get *drum roll* 224 GFlop/s. Now, obviously this isn't something that you're likely to see on a continued basis but it's there.

Now for the Xbox 360, if I had to guess, I'd have to say that each core also has a 2-issue SIMD unit, so going ahead with the same calculation you get
- 3 cores
- Each core is dual issue and is executing a FMADD
- Each FMADD is 3 mults and 2 adds per cycle
- Core clock of 3.2 GHz

96 Gflops, or roughly the 100 Gflop/s that's been floating around. So if this is correct, and it looks about right to me, the PS3 will be nearly twice as powerful as the Xbox360 CPU when operating doing dot products This is significant when you're talking about physic computations. Again, I've really skirted away the issue of keeping these SPEs fed and running at full-tilt but I think for certain datasets the Cell will destroy the Xbox360 CPU.
Essentially, when you're talking about FFT-type calculations the Cell will likely mop the floor with the Xbox360 CPU.

However, assuming that the Xbox360 architecture is more balanced than the Cell and closer to more traditional CPU design it may very well be faster in integer performance and branchy code. Now, I'm not an expert in this area, but I believe that AI code is largely branchy integer code, if this is the case, then the X360 should do well in those areas.

Again, I'm an EE student but my specialization is in analog stuff, so some of this may be off, feel free to comment and correct if I've slipped up anywhere

 

[CTho9305]

It's worth noting that your high-performance supercomputer/server/workstation is doing harder operations when rated for gigaflops than the Cell processor. Cell only gets its 200 gigaflops when using 32-bit numbers (I don't remember the exact number, but it was FAR below 100 gigaflops with 64-bit numbers), AND it doesn't provide results that are as accurate a normal system. In normal systems, floating point operations involve careful rounding of results to minimize the amount of error that builds up as you do multiple operations, but the Cell processor cuts corners here. For any scientific applications, you will want 64-bit numbers, and proper rounding of results.

The reason rounding of numbers is important is that you often use the results of one calculation as inputs to another calculation, and if you use a simple algorithm (such as "round towards 0" or "round up if >.5") after a few calculations, you'll start to see the last few digits are way off; after more calculations it really becomes significant.

Note that a GeForce 6800 can do something like 200 gigaflops too... it's easy to do lots of operations when you don't have to worry about general-purpose computation (and REALLY easy when you have a task that parallelizes well like graphics rendering). Does the PS3 have a separate graphics processor?

 

[RaynorWolfcastle]

yes, this number is for single precision floats. If you go to doubles, then you kill your throughput because I don't think that FMADD operation can be done. So basically instead of doing 5 flops per SIMD instruction you're only doing 2, at best. I remember reading somewhere that the 8 SPE Cell is only good for about 26 double-precision Gflop/s.

And yes, the GPU is good for a truckload of flops (probably more than 200 Gflops with current gen stuff since Sony is claiming performance that is "above SLI'd 6800 ultras"). Sony is claiming somewhere in the neighborhood of 1.8 Tflop/s for the PS3's CPU (known as the RSX). I'm not sure how the math works out but it's not terribly unreasonable for a 24 pipe GPU at 550 MHz, since GPUs are very similar to DSPs in that they're generally doing MAD operations. Again, correct me if I'm wrong.

 

[dguy6789]

http://xbox360.ign.com/articles/617/617951p1.html

 

[irwincur]

Now, I'm not an expert in this area, but I believe that AI code is largely branchy integer code, if this is the case, then the X360 should do well in those areas.

Most of the data I have seen seems to suggest that the Cell while being a FP beast is a slug when it comes to integer calculations. I believe that this will be the equalizer between the two designs. I suspect that performance in the real world will be similar.

 

[Schmeh]

Originally posted by: poopyman67
As soon as I came across this article, i knew something was up, so I read about the cell architecture on this site, and i thought... and i thought some more, and it just doesn't add up to me..

How does the 360 do >1tf/s w/6 cores made up of dumb downed g5s(well not necessarily dumb downed, i guess u could say "optimized" for its purpose, anways, i diverge from my point...), while a 16 core server can only(read "OMG!") 57Gf/s?

Then... they have the audacity to say that the cell is capable of 2TF/S?? What?? I read the article quite thouroughly, and with all the benefits of in-order architecture, and yada yada, I still dont see how its 20x more powerful than a 16 way server....

and @ >1/25 the cost... It seems to go against all logic and rationality....

or maybe im just plain ignorant....

the only logical explanation is that this number also includes the gpu.

thoughts?

The 2TFLOPS number that Sony released is for total system performance, not just the Cell. The same thing is true for XBox 360, 1TFLOP is the total system performance. The Cell's theoretical peak is 218GFlops, while the 360's cpu's theoretical peak is a little over 100GFLOPS.

 

[ybbor]

Good article describing what makes the xbox architecture different from a pc

Link to Article

http://arstechnica.com/articles/paedia/cpu/xbox360-1.ars

 

[imported_Sasha]

And when it comes down to it I wonder if the average gamer, playing their favorite game(s) will really care.

 

[imported_Tick]

Originally posted by: Sasha
And when it comes down to it I wonder if the average gamer, playing their favorite game(s) will really care.

HEEEEAAAATTTTTTHHHHEENNNNNNN!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!two