Post Mortem: Cell and Xenon-How good where they for uplifting game performance?

[Nesh]

I have been searching and reading threads about these two processors, there are obviously few dev documents mostly covering how some tasks were off loaded to Cell in some games, and none I could find for the Xenos.

Again these documents dont necessarily say much since a GPU with compromises like the RSX have pushed developers to find solutions by exploiting the Cell for any left over performance they could use.

Not so much is documented for Xenon. Which begs the question how much does the Xenos help in game rendering specific tasks. How much are specific tasks being optimized for it? Does/can it give a considerable amount of uplift?

Was the Cell a good choice for a console in terms of performance (excluding financial decisions)?

Given a fixed GPU (assume a 2005 model for discussion purposes) and fixed memory setup (assume 512MB unified memory) how well do those CPU's fair in a games console and how much do they help?

I also had a discussion with a computer engineer recently who is familiar with Xenon and Cell only on paper (he has never really touched any) who claims Xenon is more powerful and suitable for rendering game graphics, while Cell is suitable for math calculations (? arent graphics math anyways) and media tasks. My personal impression though was that both were kind of unsuitable for rendering graphics, Xenon was just easier to feed with general purpose tasks but less flexible in terms of programability, whereas Cell was theoretically faster, had to break down the tasks into branches to get the performance and the flexibility gave chance to developers to find tiny solutions that contribute to the graphics a bit. As I mention I am not so sure that the documents where they factor Cell's contribution to graphics is a testament of Cell's competency or simply a one way solution forced by a weak RSX


*someone correct the title. I meant Xenon not Xenos

 

[AlNom]

For reference, you might want to check out the Gamefest 2007/2008/2010 presentations dealing with a number of VMX/CPU coding issues. There are quite a lot of presentations that might be relevant to what you're looking for even if you only just check out the abstracts.

http://www.microsoftgamefest.com/pastconferences11.htm

There might be one or two presentations under the graphics section in 2010 that mentioned VMX - crowd rendering or something. I think the gist of it was that memexport was a pretty big win.

 

[mikiex]

My opinion is they should of made the RSX more powerful and had a little more memory, but then there probably would of been slower innovation with the SPUs in the CELL, then again they could of been put to less graphics oriented tasks. They didnt really get utilised in the earlier days to the extend they are now.


Xenos is a more powerful GPU, it doesnt need to offload the some of the jobs that the SPUs have been used for to help the RSX. Xenos has quirks though so its not perfect and the SPUs can do more than just support the RSX.

The designs are Apples and Oranges to me.

 

[pjbliverpool]

Anyone care to comment on how a modern desktop CPU might have faired for uplifting graphics performance? Say something like a 2500K or at the extreme end a 3960x. We all know they are insanely powerful at traditional CPU tasks but how well would they have performed as a graphical co-processor for a 2005 era GPU?

 

[Nesh]

My opinion is they should of made the RSX more powerful and had a little more memory, but then there probably would of been slower innovation with the SPUs in the CELL, then again they could of been put to less graphics oriented tasks. They didnt really get utilised in the earlier days to the extend they are now.


Xenos is a more powerful GPU, it doesnt need to offload the some of the jobs that the SPUs have been used for to help the RSX. Xenos has quirks though so its not perfect and the SPUs can do more than just support the RSX.

The designs are Apples and Oranges to me.

Its apples to oranges because you are comparing the PS3 and 360 architectures which come with more differences than just a different CPU whereas for a more like for like comparison I provided the assumptions how would they fair under a:
a) fixed GPU (preferably a 2005 model so their contribution in overall performance wouldnt be too overlapped by a relatively overpowered GPU model like the ones we have today)
b) fixed memory setup (preferably a unified 512 MB memory setup)
Under these assumptions the only differentiating factor in performance would be the CPU design between two boxes
The purpose of this thread is not to compare overall performance between the PS3 and the 360, but the CPU's themselves.

Another way to put the question is how would the PS3 and 360 performed compared to their actual counterparts if they had switched CPU's and how much does the Xenon and Cell help the GPU's in rendering graphics compared to other more commercial CPU's of their time (not in quantitative measures as its subjective and variable depending on the situation)?

 

[sebbbi]

Anyone care to comment on how a modern desktop CPU might have faired for uplifting graphics performance? Say something like a 2500K or at the extreme end a 3960x. We all know they are insanely powerful at traditional CPU tasks but how well would they have performed as a graphical co-processor for a 2005 era GPU?

AVX (in Sandy Bridge) would be very good for image post processing. AVX registers are 256 bit long, and thus a single vector can contain eight 32 bit floats, and a single instruction can process all of them (including FMA and many other complex instructions). AVX2 in Haswell can also additionally process integers in 256 bit vector registers (4 qwords, 8 dwords, 16 words or 32 bytes in a single instruction) and it has gather instructions (data loading using vector register values as array indices, a very useful feature for image processing as well).

 

[almighty]

Just to throw it in, At 5Ghz my 2500k is chucking 125Gflops around in Intel burn test while running AVX

 

[mikiex]

Its apples to oranges because you are comparing the PS3 and 360 architectures which come with more differences than just a different CPU whereas for a more like for like comparison I provided the assumptions how would they fair under a:
a) fixed GPU (preferably a 2005 model so their contribution in overall performance wouldnt be too overlapped by a relatively overpowered GPU model like the ones we have today)
b) fixed memory setup (preferably a unified 512 MB memory setup)
Under these assumptions the only differentiating factor in performance would be the CPU design between two boxes
The purpose of this thread is not to compare overall performance between the PS3 and the 360, but the CPU's themselves.

Another way to put the question is how would the PS3 and 360 performed compared to their actual counterparts if they had switched CPU's and how much does the Xenon and Cell help the GPU's in rendering graphics compared to other more commercial CPU's of their time (not in quantitative measures as its subjective and variable depending on the situation)?

It depends on what you mean by uplift, in terms of rendering what should a cpu of 2005 be doing to help rendering other than supplying ithe gpu with data to render?

 

[Shifty Geezer]

In a 2005 console, I'd say the CPU in a console should be enabling things the GPU can't do as that hasn't been invented. Things like geometry shading on Cell. Maybe post-fx as well for complex stuff that doesn't work well with SM9. One advantage of consoles over their lifespan is programmability being exploited, even using workarounds, enabling stuff that wasn't predicted when created. Admittedly that's less relevant now, but this gen had it to a degree. Certainly Kutargi saw Cell as an enable of whatever devs wanted to do, but the issues around that mean it wasn't a flawless triumph.

 

[almighty]

In a 2005 console, I'd say the CPU in a console should be enabling things the GPU can't do as that hasn't been invented. Things like geometry shading on Cell. Maybe post-fx as well for complex stuff that doesn't work well with SM9. One advantage of consoles over their lifespan is programmability being exploited, even using workarounds, enabling stuff that wasn't predicted when created. Admittedly that's less relevant now, but this gen had it to a degree. Certainly Kutargi saw Cell as an enable of whatever devs wanted to do, but the issues around that mean it wasn't a flawless triumph.

Play Station 2 greatly springs to mind, Developers over the years got that thing doing effects that it just shouldn't of been able to do...

 

[kagemaru]

Play Station 2 greatly springs to mind, Developers over the years got that thing doing effects that it just shouldn't of been able to do...

Yeah, the Saturn and PS2 were two systems that sprung to my mind reading that post.

 

[Nesh]

It depends on what you mean by uplift, in terms of rendering what should a cpu of 2005 be doing to help rendering other than supplying ithe gpu with data to render?

In terms of offloading work from the GPU giving more room for increased performance, or doing some tasks that are expected from a GPU(AA?SSAO? Deferred rendering?) or by enabling feats that were initially thought impossible by the hardware based on initial impressions and knowledge

Example: There should be something because memory constraints, RSX bottlenecks, and the SPU's limited memory led many developers and reports in 2006-2007 to suggest that the PS3 was going to be severely left behind in terms of performance compared to competition. Reading those evaluations on the hardware they sounded logical, very credible. Many saw Cell as an inappropriate, inefficient choice with lots of problems back then. All of these impressions were strongly reflected on multiform titles and uninspiring exclusive titles. Frame rate issues, lack of AA, reduction of normal maps, lower resolution textures, absence of standard effects like object motion blur (see Darkness as an example of those performance differences)......... Yet today there is a very strong parity between platforms that go against the initial beliefs. Exclusives such as Uncharted and God of War appear to defy the limitations expressed initially and squeeze out performance that was thought impossible. Exclusive games between the two platforms that are supposed to push each in unique ways are in similar levels of quality. The performance bottlenecks should have become more apparent in time as games push the envelope further and get closer to the ceiling. But this is not the case.
The Cell (and possibly Xenon) was the most unknown and unexplored component compared to all components in 2006. So it would be no strange if the CPU came with a few unexpected surprises as developers work on it and discover new stuff that it can be programmed to do
I am not sure whether this is the Cell's design to thank for or if its the possibility that they code less on the metal on the 360, leaving the Xenon's true capabilities unexplored. But both consoles push higher and higher the ceiling with 6+ year old GPU's.

 

[french toast]

wasn't the unified shaders and tesselation more rare?

Yea i read all the mumbo jumbo when both the cell and xenon were being compared and one of the 'selling points 'of the cell was flops compared to xenon, and the selling point of xenon compared to cell was 3x general processing power which was more applicable for games, where as the cell was more usefullfor media..

This was mentioned in the question, i dont suppose someone could shed some light on that?

Also, most people would say that the whole concept of the cell seemed like very advanced 'space age' stuff when announced, really good idea, but it seems something didn't quite add up, it didn't fullfill its early promise..
...why was that? L2? L3? maybe a use of edram? or was the stuff said about not being usefull for games ring true in the end?..

I would like Sony to carry the development on and see where it goes..

 

[3dilettante]

There's no L3 in Cell, no eDRAM in the PS3.

Cell's conception of multiple DSP-like cores with incompatible ISAs and separate memory spaces from a general purpose core was not new, and the objections to that layout were not new either.

What Cell brought was the integration onto a single chip and a high-bandwidth ring bus.
I haven't seen too many complaints about what was new about Cell, but many about the non-standard memory architecture and complexities in fully utilizing the more elaborate and performance-fragile architecture.

 

[jonabbey]

It's perhaps worth pointing out that Intel adopted a high speed ring bus for Sandy Bridge and Ivy Bridge as well, and we may see many more chips follow that design pattern for efficiency and flexibility.

I think whether Cell was worth it for Sony will turn on whether they decide to bring a variant of it forward for PS4. Cell may have helped the PS3 with its media handling, allowing it to support the eventual introduction of 3D Blu-Ray discs and etc., but it hasn't seemed to let the PS3 do much games-wise that the 360 couldn't do better despite releasing a year earlier.

 

[Nesh]

I think whether Cell was worth it for Sony will turn on whether they decide to bring a variant of it forward for PS4. Cell may have helped the PS3 with its media handling, allowing it to support the eventual introduction of 3D Blu-Ray discs and etc., but it hasn't seemed to let the PS3 do much games-wise that the 360 couldn't do better despite releasing a year earlier.

Well as mentioned earlier comparing the 360 and the PS3 complete architectures is not an apples to apples comparison in assessing their CPU performance. The Cell may or may have not helped the PS3 perform like it performs today. Its hard to conclude since there were other design choices in the PS3 that made things harder in general.

 

[mikiex]

In terms of offloading work from the GPU giving more room for increased performance, or doing some tasks that are expected from a GPU(AA?SSAO? Deferred rendering?) or by enabling feats that were initially thought impossible by the hardware based on initial impressions and knowledge

Example: There should be something because memory constraints, RSX bottlenecks, and the SPU's limited memory led many developers and reports in 2006-2007 to suggest that the PS3 was going to be severely left behind in terms of performance compared to competition. Reading those evaluations on the hardware they sounded logical, very credible. Many saw Cell as an inappropriate, inefficient choice with lots of problems back then. All of these impressions were strongly reflected on multiform titles and uninspiring exclusive titles. Frame rate issues, lack of AA, reduction of normal maps, lower resolution textures, absence of standard effects like object motion blur (see Darkness as an example of those performance differences)......... Yet today there is a very strong parity between platforms that go against the initial beliefs. Exclusives such as Uncharted and God of War appear to defy the limitations expressed initially and squeeze out performance that was thought impossible. Exclusive games between the two platforms that are supposed to push each in unique ways are in similar levels of quality. The performance bottlenecks should have become more apparent in time as games push the envelope further and get closer to the ceiling. But this is not the case.
The Cell (and possibly Xenon) was the most unknown and unexplored component compared to all components in 2006. So it would be no strange if the CPU came with a few unexpected surprises as developers work on it and discover new stuff that it can be programmed to do
I am not sure whether this is the Cell's design to thank for or if its the possibility that they code less on the metal on the 360, leaving the Xenon's true capabilities unexplored. But both consoles push higher and higher the ceiling with 6+ year old GPU's.

This is not just to do with Cell though, for instance lack of AA - this was always possible, it would of been not used due to ethier lack of memory or fill performance or use of a buffer not supporting MSAA. Multiple methods of getting around this were found depending on type of rendering. Again lower resolution textures is due to a lack of memory, where as now people are using smaller buffers and streaming.

My take on it, was in the begining the ps3 had a lot of potential power, but the time it took to unlock it
would be significant other than for obviously helping the RSX.

For sure the ps3 offers interesting coding experince, but still if you are doing multiplatform games its generally limited to doing things you can do on the 360 without jumping through hoops, or dont even need to worry about.

 

[french toast]

Yea i really liked the idea of Intel when they used that ring bus/L3..i really thought AMD missed a trick there....

Yea we we will see whether it was worth it for sony when the ps4 is announced.... I think what this article was getting at, was the inclusion of the SPE's worth the die space in the end?
As both a single xenon core and the cell PPE are identicle...If you were to take just one xenon core @3.2ghz and put it up against the full cell @ 3.2ghz...assuming the same gpu (RSX) and memory..(xdr2 to make it fair) what advantage would the extra SPE's actually have in a game scenario?.. (post processing that xenon/PPE couldn't do? etc)

And if it did seem to offer a tangible advantage over just a single PPE/xenos what were the associated die area costs..was it more cost effective than just going for more PPE's ala xenon?

Although im a xbox owner my self, i always liked the thinking behind the cell, if the die area costs are small enough, they could stick say 4 OoO duel thread PPE's combined to a shed load of SPE's, on that ring bus, with loads of cache on die...maybe they could use a slab of S-/edram ram as L3 unified between cell&gpu..with a slab of unified xdr2 using a big bus...

That would be very complicated and expensive, but would keep develpers going strong for years, and save Sony big in the long run as now microsoft has shown the way online, many many more $$ are there to be had over a consoles life.

 

[jonabbey]

One of the things that is explicitly open to revision in the original Cell BE architecture documents is the amount of local store in the SPUs. If Sony keeps the Cell architecture, it will be interesting to see if they decide to increase the local store, increase the number of SPUs with a fixed 256k local store, or do something more radical altogether with Cell's memory system.

If IBM produces a Power PC based CPU for Sony again, as rumor has it, I'll be surprised if there isn't at least 6 SPUs on board for backwards compatibility, though.

 

[DJ12]

Isn't ibm's latest power cpu basically cell on steroids?