Xenos - RSX - What was left out?

[SanGreal]

So noone can reconcile these two comments for me?

we don't run general-purpose code

But what you forget is, in today's game programs FP operations are 20% and the rest 80% are general integer operations or operations such as branching. They ignore that part. Integer operations are the most computation-cycle demanding part in game programs. Now, XBOX 360 has 3 times the integer processing performance of Cell.

Taken from the same interview

What is the difference between the two? Is he just talking about crap like excel in the first comment?

 

[Titanio]

With the PS3 you only have a single PPE, in-order and not very strong.

The 360 has 3 of these cores, and has much more GP power up front.

In absolute terms, though, it's weaker than desktop chips. If general purpose performance was so important, why not stick a dual core desktop derivative in there? They obviously see things more like Sony do than they're willing to admit.

In other words, dev's will be forced to rewrite code to run on the SPE's because the GP power of the PPE is insufficient.

Is it? Tim Sweeney doesn't think so. Also, what proportion of execution time is taken up by code that would be REQUIRED to run on the PPE? Tim Sweeney says: small. Let's not forget that we've got a core here running 10x faster than PS2's CPU.

(I'm sorry to harp on with Tim Sweeney this and Tim Sweeney that, and while "your mileage will vary" from game to game, he's been about the only dev to comment on this explicitly).

And again, if GP was as important as MS claims, why did they go with the setup they did?

Now, considering they already have to rewrite everything to be in-order, how much time do we expect them to have to rewrite AGAIN to optimize GP code to run using an SPE.

I'm not saying it'll be either easy or happen immediately. But if discussing technical potential..

Xenon is a little easier to work with, I've no doubt. But Cell will gain more and more as devs work harder. I think we should judge how useful those SPEs by what the best devs do with them - this is about what they can do. It's very much an open question, but it will require talent. Putting a bad driver in a ferrari and then berating its performance seems illogical - the car isn't technically any less capable than it was before.

I should also note that the PPE ISA and the SPE ISA is supposedly identical, IIRC, which should make life a little easier if trying to shift code from the PPE.

btw - i've read the document in question. It was PM'ed to me a few weeks back. I don't see how you can come to the conclusion that the GP power advantage claimed by MS were false based on that. Whatever problems the XeCPU has, the PPE on the cell will have the same issues, but with 1/3 of the power.

I'm not claiming there's no advantage, necessarily. Just that it rings a little hollow when you consider it in the bigger picture. To pull up the same analogy I used earlier, it's like an average student berating a less than average student for his academic performance - all the while completely ignoring his much greater talent elsewhere (talent which can be adapted to some unknown degree to help with his academic performance to boot).

 

[ralexand]

In absolute terms, though, it's weaker than desktop chips. If general purpose performance was so important, why not stick a dual core desktop derivative in there? They obviously see things more like Sony do than they're willing to admit.

I kind of wish that one of the console developers would have went to someone besides IBM for their cpu that way we could really see the performance difference in the in order and out of order processors.

Xenon is a little easier to work with, I've no doubt. But Cell will gain more and more as devs work harder. I think we should judge how useful those SPEs by what the best devs do with them - this is about what they can do. It's very much an open question, but it will require talent. Putting a bad driver in a ferrari and then berating its performance seems illogical - the car isn't technically any less capable than it was before.

I should also note that the PPE ISA and the SPE ISA is supposedly identical, IIRC, which should make life a little easier if trying to shift code from the PPE.

Sounds alot like what people were saying about tapping into the power of the ps2's vus vs. the celeron in the xbox.

 

[PC-Engine]

And again, if GP was as important as MS claims, why did they go with the setup they did?

What do you mean the setup they went with???

The setup as AFAICS looks like a nice balance between GP and FP.

 

[blakjedi]

How did this thread become a XeCPU vs Cell thread? I thought it was supposed to be about what was left out of the Xenos and RSX that should have been in there...

Is it because people feel that it is unfair to compare the Xenos and RSX directly (which still isnt what this thread is about mind you) because the RSX needs the Cell to operate at full capability?

Leave a thread unmonitored for a second and looks what happens. Please leave the XeCPU vs CELL argument to the billion other threads that have talked about that before.

 

[Titanio]

Sounds alot like what people were saying about tapping into the power of the ps2's vus vs. the celeron in the xbox.

There are many differences here. The VUs and the SPEs are quite different. Capability aside, many more devs will be using the SPEs more fully than those who used the VUs. PS2's main core was weak compared to the Xbox CPU also (and arguably the weak link in the whole chip). Here there is no clockspeed advantage, and the PPE and Xenon cores are reportedly pretty much the same. If you look at a pure plurality of resources - although this is somewhat simplistic admittedly - if you were to match SPEs off against cores as one might match VUs off against the Xbox CPU, that's 2 VUs (one less capable than the other) and a mips core against one celeron CPU vs one PPE against one Xenon core - pretty much the same - and 7 SPEs against the two remaining Xenon cores.

A very different kettle of fish, IMO.

What do you mean the setup they went with???

The setup as AFAICS looks like a nice balance between GP and FP.

The setup they've gone with sacrifices "general purpose" performance to try and bring them within a more respectable range of Cell-like floating point/streaming performance. If games are so heavily biased toward general performance as they suggest (80/20), why did they do this?

 

[PC-Engine]

The setup they've gone with sacrifices "general purpose" performance to try and bring them within a more respectable range of Cell-like floating point/streaming performance. If games are so heavily biased toward general performance as they suggest (80/20), why did they do this?

Maybe because they couldn't put 3 G5 cores with 3 VMX units on a single die?

 

[scooby_dooby]

I don't know, I think if you judge by the "best" possible, the CELL wins everytime. I think that's a ratehr convenient way of looking at it if you are a CELL fan.

What about a more realistic outlook? Like, what really will the SPE's be able to do, how un-efficient will they be USUALLY, what's the likelihood of extracting all this power?

As for the PR comments, hey sony plays up FLOPS, so MS plays up GP, does it really matter? Anything both camps say is pretty much tongue-in-cheek, like does anyone actually believe that they believe what they're saying?

I don't see the relevance of comparing to a desktop CPU, there's many other reasons not to choose a desktop CPU, mainly costs. In comparison to the PS3, which is the only comparison that matters, it does have very strong GP performance.

As for their claims ringning hollow I guess it depends how you look at it. You might argue MS considered the lack of GP power in the PS3 a large design flaw, and that MS was simply pointing out a flawed design, and although their CPU is no GP moster, it certainly has much more power than the PS3 does, so what';s wrong with them pointing out this rather glaring weakness?

At the same time I don't know if Sony can make the same claim in regards to X360 and it's FLOP power, as it is quite strong and definately couldn't be described as weak.

i think MS believes they have a better balanced system, and the comments referring to GP power were to try and illustrate that point. Take the focus off FLOPS for a little, and put the spotlight on the PS3's big weakspot, GP performance.

 

[Titanio]

Maybe because they couldn't put 3 G5 cores with 3 VMX units on a single die?

But even just 2 would apparently be better for GP? And that's your focus in a game...

...right?

And they wouldn't have to be G5s to be more GP.

 

[pegisys]

The setup they've gone with sacrifices "general purpose" performance to try and bring them within a more respectable range of Cell-like floating point/streaming performance. If games are so heavily biased toward general performance as they suggest (80/20), why did they do this?

because they wanted games to be easy to port, they know sony is still going to have more support, even if it doesn't stay that way, and if games are made with high floating point performance in mind they don't want to be waiting long periods of time for a port, I think that may have been the reason they went with ppc in the first place instead of a multicore x86 cpu, well that and maybe price

 

[PC-Engine]

Maybe because they couldn't put 3 G5 cores with 3 VMX units on a single die?

But even just 2 would apparently be better for GP? And that's your focus in a game...

...right?

And they wouldn't have to be G5s to be more GP.

Maybe FP wasn't high enough with only 2 G5s? Why not just use one core with a 80/20 split. See what I'm getting at? If their minimum was 60 for FP then they would need 240 for GP which wouldn't fit into a single die so instead of cutting the 60 which is the bare minimum, they went and cut the 240 down to 180 to fit on a single die.

 

[version]

G5 isnt SMT, x360's large memory latency kill the g5's performance

 

[overclocked]

Dont know how the Xcpu differences from the PPE but i think that many here forget bacause they are so biased that $ony have had quite some more time with its partners to deliver Cell vs MS and even that is an understatement. Although all the cash and time spended on Cell wont "guarantee" a sucess i find it likely that Cell is indeed going to deliver much in terms of performance. Compared to PS2 that was a pain to code for and compare to Cell witch already have tools and software this long before launch.
So i think the ones here that dont belive in Cell as a good Cpu will change your mind when your in the local compstore next year.

Btw all this bias and fanboyism is really hurting the Forums and although Sonic as a mod keeps up a good spirit most of the times i really think that you Jvd should hold your negatives against $ony because you are a mod and its not really serving this commuity with an attitude like that. Keep it for yourself or quit as a mod and let someone else less biased take your role.

 

[Titanio]

As for the PR comments, hey sony plays up FLOPS, so MS plays up GP, does it really matter?

Yes, it does. People often say this - Sony played the flops game, so why should we complain about MS? Well, it IS different! As already mentioned earlier, Sony used a well understood and verifiable metric (on paper), and credited Xenon fully. MS's "metric" is completely non-transparent and completely arbitrary. As long as the SPE's potential contribution to "general purpose" performance is non-zero, that comparison is flawed. Furthermore, it's very annoying, because while it's "case closed" on Cell's biggest strength (fp), we remain in this quagmire of conclusion-less debate about an area of supposed lesser performance for Cell, that may or may not even be that big an issue (see: Sweeney). Pretty clever on MS's part - because we can't come to any conclusion, this debate about a Cell "weakness" will continue - and that's just how MS likes it.

because they wanted games to be easy to port, they know sony is still going to have more support, even if it doesn't stay that way, and if games are made with high floating point performance in mind they don't want to be waiting long periods of time for a port, I think that may have been the reason they went with ppc in the first place instead of a multicore x86 cpu, well that and maybe price

Well, porting from PC isn't going to be very easy - some devs are crying for their intels and athlons, it seems. And if they want to be closer to Sony, why not go the whole hog? I'm not so sure that they didn't because they didn't want to.

Maybe FP wasn't high enough with only 2 G5s? Why not just use one core with a 80/20 split. See what I'm getting at? If their minimum was 60 for FP then they would need 240 for GP which wouldn't fit into a single die so instead of cutting the 60 which is the bare minimum, they went and cut the 240 down to 180 to fit on a single die.

60 for FP? What?

If things were as biased toward GP as they claim, why didn't they bias their chip toward GP instead of this "balance"..?

 

[scooby_dooby]

It seems like you think MS invented the idea that SPE's are horrible for GP code.

Has this not always been pretty much accepted? They are floating point monsters that's their strength, and from what I can gather not only are they weak at integer operations they are qctually quite terrible.

Now I understand your point, that the comparison is flawed. However, what do you propse MS should have done?

Do you think they should simply have not made the comparison at all? Despite that being their MAIN CPU advatange of the CELL? let Sony win the day with their FLOP count and stay completely silent about their GP advantage?

Or should they have come up with some arbitrary, unproven, value for each SPE's GP performance?

I mean I don't see really how they could make a comparison and accurately include the GP performance of teh SPE's.

But honestly, they are supposed to be pretty terrible, so what's the likelihood of them being used for anything signifigantr in real world scenarios?? Given that, I think the comparison is more or less valid even if it isn't the magnitude they suggest.

 

[pegisys]

Well, porting from PC isn't going to be very easy - some devs are crying for their intels and athlons, it seems. And if they want to be closer to Sony, why not go the whole hog? I'm not so sure that they didn't because they didn't want to.

that might be as close as they can get to the cell without just using one themselves, I don't know if sont would have allowed that, and I don't think they gave pc ports any thought in hardware design, thats what they have directx and XNA for

 

[PC-Engine]

60 for FP? What?

If things were as biased toward GP as they claim, why didn't they bias their chip toward GP instead of this "balance"..?

The ratio 80/20 x 3 = 240/60. If 60 was the bare minimum then the ideal would have to be 240 for GP. Since 240 could not be achieved in the die size limit, they had to decrease that to 180 or around there. They could not decrease the 60 any further since it was the bare minimum.

 

[scooby_dooby]

The ratio 80/20 x 3 = 240/60. If 60 was the bare minimum then the ideal would have to be 240 for GP. Since 240 could not be achieved in the die size limit, they had to decrease that to 180 or around there. They could not decrease the 60 any further since it was the bare minimum.

I don't think you really believe your own argument. How could you? Xe360 has over 100GFLOPS, that's some serious FLOP power compared to today's CPU's so they obviously invested heavily in floating point performance.

Allard's statements shouldn't be taken verbatim, they are meant for marketing purposes, to make headlines and to downplay their competition. They say GP is the most important, but at the same time they create a very powerful floating point CPU.

What's more likely to be a true indication of their priorites? What they say? Or what they do?

What they actually DID was create a very balanced system with strong FLOP and GP power.

 

[Titanio]

I mean I don't see really how they could make a comparison and accurately include the GP performance of teh SPE's. But honestly, they are supposed to be pretty terrible, so what's the likelihood of them being used for anything signifigantr in real world scenarios??

Err..you mean anything significant in terms of "General purpose" processing?

Do you think the contribution will be zero? I don't think you do. My point is if they even only offered a tenth of the performance combined as the PPE, that's enough to render the comparison flawed (and I see you accept that).

IMHO, they shouldn't have made the comparison at all. Why did they even feel the need to try and compare? They could have stayed quiet on this front and pushed the whole package, as they seemed to kind-of be interested in doing (it's all about services etc. etc.) but then they just couldn't resist trying to cobble together a technical defense simultaneously. They made such an issue of GP performance, were at pains to point out its importance to games and to throw out numbers as some sort of backup. But when it emerges that your strength isn't so strong in absolute terms...well that's the sound of your argument deflating. And the question still remains as to why they didn't build their chip to match what they're claiming about game behaviour instead of this "half-way house" - the argument does ring very hollow IMO.

The ratio 80/20 x 3 = 240/60. If 60 was the bare minimum then the ideal would have to be 240 for GP. Since 240 could not be achieved in the die size limit, they had to decrease that to 180 or around there. They could not decrease the 60 any further since it was the bare minimum.

I see your ratio (not that I think it's actually representative), but I don't know where your other numbers are coming from. Xenon does not appear to have been built around this ratio.

 

[PC-Engine]

I see your ratio, but I don't know where your other numbers are coming from. Xenon does not appear to have been built around this ratio.

I think they added as much FP power as a precaution since they don't know what future game engines will need.

Anyway that ratio was just to show why GP didn't end up as 80/20 but something else instead. That something else may be just as a precaution.