New interview about Ps3 CM:Dirt

[pjbliverpool]

Why would you expect Xenon to be significantly behind a high end PC? You've got three large vector engines in there at 3.2 GHz, and a closed hardware system where cache and data management can help keep things more efficient.

Aside from the vector engines each Xenon core has about 1/3rd the execution units of an Athlon64 (single core) and then there is the obvious lack of OOOE. It was my understanding that the PPE's were widely accepted as being much slower than a modern PC core.

On paper Xenon pushes out about the same raw GFLOPs as a QX6800 but I don't buy for a moment that it could ever match it in.... well anything!

EDIT: Additionally, isn't the PPE supposed to be pretty bad at branchy code anyway? The combination of a deep, narrow pipeline, small cache and minimal branch prediction logic is a perfect recipe for poor performance in that type of code. I wouldn;t be suprised if all 3 cores in combination couldn't keep up with a single core Core2 at a modest clock speed.

 

[Darkon]

Aside from the vector engines each Xenon core has about 1/3rd the execution units of an Athlon64 (single core) and then there is the obvious lack of OOOE. It was my understanding that the PPE's were widely accepted as being much slower than a modern PC core.

On paper Xenon pushes out about the same raw GFLOPs as a QX6800 but I don't buy for a moment that it could ever match it in.... well anything!

It's a question of whether you'd really consider shuffle to be "FLOPs", MS was counting VMX FMADD (8) + VMX shuffle/permute (4) together if i remember correctly

 

[pjbliverpool]

It's a question of whether you'd really consider shuffle to be "FLOPs", MS was counting VMX FMADD (8) + VMX shuffle/permute (4) together if i remember correctly

I guess not since the Intel numbers probably don't.

Anyway its either 115.2 or 76.8 vs 93.8.

 

[Shifty Geezer]

Aside from the vector engines each Xenon core has about 1/3rd the execution units of an Athlon64 (single core) and then there is the obvious lack of OOOE.

Things like OOOe, Cell's SPUs lack too, but they mince through Physics quite nicely.

It was my understanding that the PPE's were widely accepted as being much slower than a modern PC core.

Overall, sure. But 1) Xenon's cores != Cell's PPU, and 2) if you can run the physics on the VMX units, there's no worry. These units aren't cut down AFAIK, indeed having more registers and more room for work. I don't know how a physics engine is designed to run on a VMX enabled core, and if it requires a lot of input from the non-vector maths units, that would be an issue. But I expect a lot of a physics engine can be directed to these units.

On paper Xenon pushes out about the same raw GFLOPs as a QX6800 but I don't buy for a moment that it could ever match it in.... well anything!

Peak flops show how fast you can run without data and other bottlenecks. The actual peak (ignoring different definitions of what a flops is!) is there if you can provide the data efficiently enough. If Xenon's cache management is competent enough, it ought to be good in that department. No processor will attain its peak in real use, but the measure is how close you can get. Xenon should be more pushable than a PC CPU in that respect.

EDIT: Additionally, isn't the PPE supposed to be pretty bad at branchy code anyway? The combination of a deep, narrow pipeline, small cache and minimal branch prediction logic is a perfect recipe for poor performance in that type of code. I wouldn;t be suprised if all 3 cores in combination couldn't keep up with a single core Core2 at a modest clock speed.

Yes, but that's the same as the SPU's. If your physics engine works to the strength of the processor, massive FP throughput (like the Aegia PPU), then fast, big vector units == strong physics. Basically the things that make SPU's good at physics apply to some degree for Xenon. It might not be as fast as a quad-core whatnot, or maybe even as fast as a Dualcore, but Xenon shouldn't be way behind, if it is behind.

 

[pjbliverpool]

Yes, but that's the same as the SPU's. If your physics engine works to the strength of the processor, massive FP throughput (like the Aegia PPU), then fast, big vector units == strong physics. Basically the things that make SPU's good at physics apply to some degree for Xenon. It might not be as fast as a quad-core whatnot, or maybe even as fast as a Dualcore, but Xenon shouldn't be way behind, if it is behind.

Fair enough but the local store, main memory bandwidth and the fact that there are 7 of them likely have just as much impact, if not moreso on physics that the raw floating point power in Cell. Remember because of the LS the SPU's have 50% more memory to play with than Xenon per thread (not even cosidering the PPU), and its cache size which could be one of the main sticking points for physics in Xenon.

Also I may have just missed it but I havn't really seen any definitive proof of Cells superiority for phyiscs in in game situations. Lots of talk about what it *should* be able to do lots of scientific benchmarks with questionable applicability but nothing relating purely to physics as it occurs within a game. Like I say though, I may have just missed it.

 

[Arwin]

Fair enough but the local store, main memory bandwidth and the fact that there are 7 of them likely have just as much impact, if not moreso on physics that the raw floating point power in Cell. Remember because of the LS the SPU's have 50% more memory to play with than Xenon per thread (not even cosidering the PPU), and its cache size which could be one of the main sticking points for physics in Xenon.

Yeah. An important part of the first discussions of the Cell by sites like Anandtech was that the special thing about the Cell's SPU's is that it seems it is possible to actually achieve the theoretical performance by virtue of the way the Flexio, Local Stores and SPE memory management is set up.

Also I may have just missed it but I havn't really seen any definitive proof of Cells superiority for phyiscs in in game situations. Lots of talk about what it *should* be able to do lots of scientific benchmarks with questionable applicability but nothing relating purely to physics as it occurs within a game. Like I say though, I may have just missed it.

Havok's talk at I think an IBM conference about how they optimised their physics engine, which is widely used in games, for the SPUs and achieved a really big performance improvement, more than they themselves expected. Here's a reference to that Havok 4.5 announcement on Joystiq, quoting an improvement of factor 5-10 and mentioning it was done in cooperation with the Motorstorm guys:

http://www.joystiq.com/2007/01/25/new-havok-release-supports-ps3/

 

[pakotlar]

Thats news to me! Especially in terms of the 360. If anything I would expect the 360 to be significantly behind a high end PC in the physics departement (without considering the impact of a PPU).

X360 CPU is actually supposed to be very good at physics based code. It can handle what, ~100 Gflops for trivial workloads? I know that it was revised from around 75Gflops, so its somewhere in that range. A C2D or A64 are nowhere near that #. Although a quad core cpu is probably close in that department. Remember we are talking strictly about physics, i'm assuming mainly rigid body stuff, collision detection...

 

[Darkon]

X360 CPU is actually supposed to be very good at physics based code. It can handle what, ~100 Gflops for trivial workloads? I know that it was revised from around 75Gflops, so its somewhere in that range. A C2D or A64 are nowhere near that #. Although a quad core cpu is probably close in that department. Remember we are talking strictly about physics, i'm assuming mainly rigid body stuff, collision detection...

Hard to believe after seeing cell's PPE performance

 

[pjbliverpool]

X360 CPU is actually supposed to be very good at physics based code. It can handle what, ~100 Gflops for trivial workloads? I know that it was revised from around 75Gflops, so its somewhere in that range. A C2D or A64 are nowhere near that #. Although a quad core cpu is probably close in that department. Remember we are talking strictly about physics, i'm assuming mainly rigid body stuff, collision detection...

Im no dev so what I know about this is only from reading articles on the web but its my understanding that their is a lot more to calculating physics that just floating point power. Integer throughput, memory bandwidth and perhaps most importantly, branch prediction all have a large impact (as I understand it).

According to Arstechnicas article on Xenon, its deep, narrow pipeline(s) small cache and presumed poor branch prediction logic should seriously hamper physics (and AI) performance.

Regarding the floating point comparisons, a QX6800 comes in at 93.8 GFLOPS while Xenon measured in a comparable manner comes in at 76.8 GFLOPS, so yeah, they are in the same ball park there. However when you consider the Intels massive Integer, Cache and branch prediction advantages, plus the fact thats its OoO, I would have thought it would easily exceed Xenons performance in physics.

 

[rekator]

Hard to believe after seeing cell's PPE performance

A Xenon core is not the PPE in the CELL…
There's both based on a PowerPC design but they're not the same… like a G4 or G5…

And for performances of the VMX in Xenon it's difficult to have a goog idea because It's the first PowerPC (at least for general purpose) who got a VMX unit with 128 registers… (G5 got 64 registers and CELL PPE 32…) and got some specifations (and lack over compare to G5's VMX) so we can make direct interpretation from other VMX PowerPC Unit, only supositions .

For the In-Order execution, I think it's less penalysing for a console CPU than a computer CPU… It's a close box where only Console OS and a unique tailor-made game runs, not like in a computer where plainty of differents non well optimised programs runs.

 

[sevanig]

Being very non-technical. After playing most of the 360's games since launch, even the new flatout game where the collision detection is not good + floaty physics...
And then after playing the ps3, specifically motorstorm. you can see the ps3 is better at physics calculations.

Both games being exclusive to each platform, and motorstorm being a 1st generation ps3 game (i think latest version of havok was released after motorstorm shipped).

I look forward to see what more optimization can do (motorstorm 2)...

 

[Shifty Geezer]

Im no dev so what I know about this is only from reading articles on the web but its my understanding that their is a lot more to calculating physics that just floating point power. Integer throughput, memory bandwidth and perhaps most importantly, branch prediction all have a large impact (as I understand it).

All of that depends on the engine. One only needs look at the optimizations produced in 'SPE unfriendly applications' that makes it run well on Cell, to see how entrenched ideas about problems can be overcome by casting them aside and targetting the hardware.

According to Arstechnicas article on Xenon, its deep, narrow pipeline(s) small cache and presumed poor branch prediction logic should seriously hamper physics (and AI) performance.

I think Ars is as entrenched in old preconceptions as many who have refuted the importance are fast vector processing. Again, stuff that Cell is no good at by design, it's doing very well at, because the algorithms have been redesigned. You couldn't target super-vector physics engines for CPU's that were better at branching and hiding memory access function from the developer, so the engines were built around those CPU designs. A stream-processor targetted engine will throw branching to the wind and handle things differently to make the most of the processors.

 

[Mintmaster]

And then after playing the ps3, specifically motorstorm. you can see the ps3 is better at physics calculations.

Both games being exclusive to each platform,

That is a very naive statement. Motorstorm's physics has nothing to do with the hardware's capabilities. It has everything to do with the code written by Evolution Studios.

Moreover, the thought that every developer of exclusive games is equal in talent is absolutely absurd.

 

[Arwin]

That is a very naive statement. Motorstorm's physics has nothing to do with the hardware's capabilities. It has everything to do with the code written by Evolution Studios.

Moreover, the thought that every developer of exclusive games is equal in talent is absolutely absurd.

Now you yourself are being naive. Motorstorm's physics are driven by Havok's multi-platform physics engine, and unless you've since that announcement learnt something I haven't, Havok themselves have indicated that as of Havok 4.5's optimisations, the Cell gives by far the best performance here and that they feel this is because the Cell's SPU based architecture works exceptionally well for physics calculations.

 

[Mintmaster]

Now you yourself are being naive.

Excuse me? How exactly am I being naive? Nothing in your post is relevant to anything I said.

Motorstorm has nice looking physics because the devs made models with lots hinges/joints connecting parts together that break spectacularly during crashes. FlatOut doesn't even use Havok as far as I can see.

So we have a self-admitted non-technical gamer subjectively judging the physics between two different games from two different dev studios using two different physics engines. From that we can conclude PS3 is much better at physics than XB360?

Get real.


I guarantee you that everything Motorstorm does is doable on XB360, but obviously I can't prove it. If LittleBigPlanet was multiplatform and you created a level where you stacked craploads of things on top of each other to the point where the game slowed down on the 360 and not the PS3, then you have possible proof that PS3 is better at physics (you still have to profile the game to be sure that nothing else is limiting performance).

Until then, there is no definative answer. Even your "proof" of some qualitative statement from Havok is almost irrelevant. In game scenarios vary tremendously, especially when you take into account system loads due to the GPU. I could even make an argument that extensive, immersive physics interaction in a 3D world is limited by RAM space. Fully destructable environments, for example, would require tons of space to store the state of every wall in a game level.

 

[Arwin]

Excuse me? How exactly am I being naive? Nothing in your post is relevant to anything I said.

Yes it is. You were saying that Motorstorm's physics had nothing to do with the hardware and everything with the code they had written. I'm not saying sevanig isn't being naive here. But your comment shoots off so much in the other direction that it's not much better, now is it?

Although I expect you now probably see my point, just to mock you a bit I'm going to continue this in detail.

Motorstorm has nice looking physics because the devs made models with lots hinges/joints connecting parts together that break spectacularly during crashes.

So that's all it takes to make a nice looking physics based game huh?

FlatOut doesn't even use Havok as far as I can see.

Don't make a habit out of this 'as far as I can see' thing. Googling for 'FlatOut and Havok' is but a small effort and saves you from making a fool out of yourself.

So we have a self-admitted non-technical gamer subjectively judging the physics between two different games from two different dev studios using two different physics engines. From that we can conclude PS3 is much better at physics than XB360?

So now we have the sentence 'now you yourself are being naive'. From that we can conclude that I didn't think sevanig was being a little naive also? But he thinks he is starting to see a pattern, and he has every right, especially since he qualified his comments as coming from a consumer point of view, to do so. Then we have every right to point out that current games may not be pushing physics all that hard, and/or the actual physics calculations may or may not be the limiting factor.

You can argue all you like about different situations though, but let's just say that I personally think that the PS3's is better at physics than the 360's CPU, and I think there may even be a fair bit of consensus about that in the community. Whether or not the difference in capacity actually matters in the day to day practice of gaming, and whether or not the 360 can then compensate for this by virtue of having a better/more flexible GPU, who knows? Clearly both consoles have different strengths and weaknesses.

We can discuss this for quite a bit and go into a lot of technical details, but people like sevanig are going to judge this simply by looking at the games. And it'll probably be just as interesting to see if we will be able to spot a difference there as whether or not we can conclusively prove with cache, latency figures and what not which hardware is best at what in theory.

Havok is used in a ***load of games though, even previous-gen, so that it runs above average on Cell is certainly good news for PS3 owners and it is not that far-fetched to assume that multi-platform games may benefit from it at some point. I personally think we'll find out soon enough.

And yes, I agree with you that probably the 360 can do Motorstorm physics, and I defintely think it can do LBP physics. The platforms are still young anyway, so there's typically a lot of room for growth.

Sevanig, I think that the Havok 4.5 engine was partly a result of working with the Motorstorm guys to optimise their engine for Cell. So Motorstorm may well have a number of those optimisations already in there, if not all.

 

[ErnstH]

You can argue all you like about different situations though, but let's just say that I personally think that the PS3's is better at physics than the 360's CPU, and I think there may even be a fair bit of consensus about that in the community. Whether or not the difference in capacity actually matters in the day to day practice of gaming, and whether or not the 360 can then compensate for this by virtue of having a better/more flexible GPU, who knows? Clearly both consoles have different strengths and weaknesses.

Even if the Cell is better suited for physics, it's not so much better than Xenon to actually make a visible (or meaningful) difference. And that's what is relevant.
(although this works both ways)

 

[Arwin]

Even if the Cell is better suited for physics, it's not so much better than Xenon to actually make a visible (or meaningful) difference. And that's what is relevant.
(although this works both ways)

Like I said, we'll see.

 

[sevanig]

Thanks for clearing that up Arwin.
I thought starting my sentence by saying "Being very non-technical" would save me from being flamebait.
But even after owning motorstorm for 2 months, I still see new situations in the crash slow-mo's (even the AI players crashing in front of you), and still I am amazed by the amount of collision detection/physics going on (more so the collision detection of all the car parts with the environment/cars/mud).
I have never really seen anything like it before (except for those ageia physics card demos and two of the cut-scenes in gears of war)

I cant wait to see what else they can come up with...

 

[Teasy]

Thanks for clearing that up Arwin.
I thought starting my sentence by saying "Being very non-technical" would save me from being flamebait.

To be honest if you're very non-technical then you probably shouldn't be making firm statements about technical issues No offence..