Cell's PPE and Xenon compared.

[ADEX]

Compared the Cell's PPE and a 360 core:

The colours are different and the Xenon's core image is somewhat blurry but look closely and you can see the same structures. The integer, FP and control sections are all the same, the vector sections however are quite different. You can clearly see the additional registers in the Xenon (128 Vs 32 in the PPE). There are also additional structures in the Xenon which are not present in the PPE, these'll be for the additional instructions no doubt.

The Xenon core also appears to be very slightly taller then the PPE by a fraction of a millimetre.

For the full Xenon die size I'm estimating 154 sq mm at 90nm.

 

[scificube]

Just by observation it would appear the PPE is wider than the X2 core by approximately same amount as it is shorter than the X2 core going by the orientation of your pics. I've no clue as to how you scaled things so that's about as precise as I can get. This may suggest that the location or orientation of something(s) has changed but are still common to both cores.

I suppose a good question to ask is what logic is now consuming the space in the PPE that the larger register pool takes up in the X2 core? I don't think anyone has a good answer just yet...if so it be interesting to hear it. Of course I am assuming the space wasn't simply left empty or consumed by arranging things differently than in the X2 core.

 

[Nemo80]

The Xenon Core is very much zoomed in the above picture. As others have shown, the Xenon Core is more similar to a DD1 CELL PPE, which is about half the size of the final DD2 PPE.

or look at this one:

 

[Lysander]

half the size! what are you talking about?

 

[Nemo80]

half the size! what are you talking about?

The DD1 PPE is about half the size of the DD2 PPE (which is used in PS3). The DD2 is most likely similar to a dual core SMP, sharing very little resources.

 

[Lysander]

There is no similarity between X2core and dd1 at first sight. The dd1 pic looks a bit weird, too narrow. You say than the main core of cell is "dual-core"?!

 

[Jawed]

The Xenon VMX FP pipeline (vector scalar unit) is about twice as long as Cell PPE's. As far as I can tell this is because of Xenon's Dot Product pipeline, which is the longest of all pipelines.

Jawed

 

[version]

XENON's VMX floating simd unit longer than PPS's, because dotproduct,DX compression instructions etc

PPS's vmx contain 8 floating units, but i dont know why,
it has more facilities:

1. 2*4 ways floating simd unit
2. 1*8 ways simd
3. yield problem or higher frequency
4. VMX run at half speed

 

[ADEX]

Just by observation it would appear the PPE is wider than the X2 core by approximately same amount as it is shorter than the X2 core going by the orientation of your pics. I've no clue as to how you scaled things so that's about as precise as I can get.

That's an optical illusion. I initially scaled them to the same width and height but had to increase the height of the Xenon as the structures didn't quite line up.

They don't look the same on first impression but I zoomed in and looked part by part, the colours are different and there is distortion due to the lack of resolution but if you look close enough the shapes of many of the structures are the same.

It looks like IBM have created a new core (the mythical PowerPC 350/360) then used it unaltered in the Cell and modified it for the Xenon. Makes a lot more sense than creating a completely new core for each.

 

[scificube]

I figured as much. Sorry been busy today so I didn't respond sooner.

Jawed has some nice info about the VMX unit. It appears we need to look elsewhere in order to give a context to what Crytek has recently said..or find some collarboration of their thinking about the PPE in contrast to an X2 core. It could be just their take on things.

 

[Guden Oden]

They don't look the same on first impression but I zoomed in and looked part by part, the colours are different

Actually, there aren't any colors at all really... What you see in die-shots like these are chips photographed under polarized light, so depending on the angle of the light/polarization filter, the colors will change. Probably dramatically as well.

 

[Megadrive1988]

sorry I've forgotten, but what is the likelyhood that PS3 is getting the DD2 version of Cell ?

 

[Acert93]

sorry I've forgotten, but what is the likelyhood that PS3 is getting the DD2 version of Cell ?

I believe it is confirmed that the PS3 will be getting the DD2 CELL revision.

 

[mboeller]

I believe it is confirmed that the PS3 will be getting the DD2 CELL revision.

IMHO Strange;

the PPC970 is only 33% larger than the DD2 so why didn't they switch directly to an PPC970 core? The die-area would not get much bigger due to the small size of the PPC-core compared with the complete core. This would help the PS3 a lot cause the performance should be higher and it should be much simpler to program the CPU due to the OOO PPC970.

 

[Panajev2001a]

IMHO Strange;

the PPC970 is only 33% larger than the DD2 so why didn't they switch directly to an PPC970 core? The die-area would not get much bigger due to the small size of the PPC-core compared with the complete core. This would help the PS3 a lot cause the performance should be higher and it should be much simpler to program the CPU due to the OOO PPC970.

"only 33% larger" ?!? I do not think that noone wanted the PPE core to be even 1'' bigger than it is with DD2: the transition from DD1 PPE to DD2 PPE must have been already obect of many discussions and badly needed. Augmenting the whole CPU by smething like 30% extra seems a bit too much.

 

[LunchBox]

Compared the Cell's PPE and a 360 core:

The colours are different and the Xenon's core image is somewhat blurry but look closely and you can see the same structures. The integer, FP and control sections are all the same, the vector sections however are quite different. You can clearly see the additional registers in the Xenon (128 Vs 32 in the PPE). There are also additional structures in the Xenon which are not present in the PPE, these'll be for the additional instructions no doubt.

The Xenon core also appears to be very slightly taller then the PPE by a fraction of a millimetre.

For the full Xenon die size I'm estimating 154 sq mm at 90nm.

the Xenon core pic looks like it's stretched a little bit.

not a properly scaled picture for comparison.

 

[Nemo80]

the Xenon core pic looks like it's stretched a little bit.

not a properly scaled picture for comparison.

As i said, the Xenon core is much like the DD1 PPE which contains also only one VMX unit. The DD2 PPE is almost "dual-core" like (not SMT like Xenon, but SMP like), containing a second VMX unit and additional logic which the Xenon lacks. Guess this is because Sony initially wanted to have two DD1 Cores, but instead they got 1 "dual-core".

 

[LunchBox]

As i said, the Xenon core is much like the DD1 PPE which contains also only one VMX unit. The DD2 PPE is almost "dual-core" like (not SMT like Xenon, but SMP like), containing a second VMX unit and additional logic which the Xenon lacks. Guess this is because Sony initially wanted to have two DD1 Cores, but instead they got 1 "dual-core".

I see.

Sorry...

I only saw the first post and roughly skimmed through the replies.

thankies for pointing that out.

 

[Jawed]

http://www.beyond3d.com/forum/showpost.php?p=561024&postcount=148

Jawed

 

[3N1gM4]

Both the PPE and XCPU will probably have 2 VMX units. Having just one unit would be too much of a backward step in terms of performance.

More likely, the SIMD integer unit has been scraped in the XCPU to accommodate the longer floating point op-codes and the larger register file that they use. It’s possible that this may have been the only feasible option available, without having to resort to register banks, due to the lack of sufficient op-code space.