RSX: 1.1 billion vertices/sec?

[ERP]

And I heard that some XBox games go as high as 50M PPS textured and lit.

I don't think there's a game that came even close to that. The highest I've ever heard was about half of that figure.

Also, for car games, Forza, RSC2 and other top XBox games are budgeted like 15,000~20,000 polys per car while I think GT4 is budgeted around 5000~6000 polys. That's why top XBox car games usually have car damage for instance.

There are games on PS2 that have a lot more polys on cars than GT4. I forgot the numbers but I think resident devs can chime in on this. I think that one F1 game in particular was said to push around 18Mp/s by it's developers. Also, consider most of the Xbox racing games run at 30FPS (except for RSC/2) which effectively halves their poly/sec throughput. Frankly, I can't think of a game where you'd actually have 3x more polys/sec on Xbox than on a comparable PS2 game. More yes, but not nearly that much more.

Oh Dear god please save me from 5 years of PS2 was better at pushing polys than XBox threads. Were only now starting to see less Saturn was better than PS threads.

This is pretty simple, on every single application I have ever tested Xbox is faster than PS2 (sometimes by enormous margins) with one notable excepion, when the games become transparent fill limited (loadsa particles) then Xbox is slower.

But then you'd expect that it shipped 18 months later. But PS2 has the EDRAM which explains the aberation.

I have never heard of anyone building a cross platform game worrying too much about the Xbox version, the same devs often have to invest a huge amount of time to get the PS2 version up to speed.

 

[London Geezer]

Not per frame, but per second.

Game that renders say 1000 polygons per frame, will display 30000 polys/sec if a game is running at 30FPS game, and 60000 polys/sec if the game is running at 60FPS. That's what I meant.

Same thing i was saying, just from the other side.

When going from 30 to 60 fps, when polys per frame budget is constant, you'd have to push double the polys per second.

When polys per second is constant, going from 30 to 60 will halve your polys per frame budget.

 

[Agisthos]

Why are people still debating these gpu numbers as a measure of performance? I thought it was long past the time of realising the theoretical max numbers mean nothing.

Its like seeing 800x600 benchmarks with no AA, who cares.
If a developer has a game running at 60fps with full effects on and 4xAA then it would be cool to hear of the polygon throughput of that in comparison to last gen's games.

 

[gotgame]

With the difference that those games also push AA and real-time shadows.Bizarre stated a few times that the could get 60fps with no AA but rather have the AA.Even with the AA it ran at higher than 30fps but it wasn't stable.

Sometimes I wonder if this board really is more "fanboy" proof than a normal gaming board.

 

[Qroach]

There are games on PS2 that have a lot more polys on cars than GT4. I forgot the numbers but I think resident devs can chime in on this. I think that one F1 game in particular was said to push around 18Mp/s by it's developers. Also, consider most of the Xbox racing games run at 30FPS (except for RSC/2) which effectively halves their poly/sec throughput. Frankly, I can't think of a game where you'd actually have 3x more polys/sec on Xbox than on a comparable PS2 game. More yes, but not nearly that much more.

I can't beleive this is till being argued. There's lot's of developers that say otherwise and peopel STILL don't listen. it's freaking hopeless.

 

[j^aws]

Let's see...

Assuming ~1 vert per triangle. 1 triangle set up per clock. The vertex shaders need to produce one vertex per clock to keep up. This means a maximum of 48 Vec4+Scalar ops per vertex. A more well balanced scenario between vertex and pixel shading might mean 12 Vec4+Scalar ops per vertex. Not all that fancy, but also not trivial. The 500 million polygons per second number of the 360 looks a lot more reasonable than the 125 million number thrown around when the original XBox was announced, or the 67 million number for the PS2.

On the RSX side, running the same vertex shader...
12 ops (Vec4+Scalar) per shaded vertex gives us, 8 * 550 / 12 = 367 million polygons per second.

This is where the flexibility of CELL SPUs may come in handy.

VS unit ~ vec4 + scalar ~ 10 flops/cycle*0.55gHz ~ 5.5 GFlops

SPU ~ 8 Flops/cycle * 3.2 GHz ~ 25.6 GFlops

Depending on the CELL<=>RSX integration, using a couple of SPUs + VS units, you could push it closer to 1 Billion poly/sec I guess...and you'd still have those powerfull PS units for fragment shading. Furthermore, another SPU could be used for post-proccessing. So essentially the SPUs would load-balance with RSX...

 

[ERP]

Let's see...

Assuming ~1 vert per triangle. 1 triangle set up per clock. The vertex shaders need to produce one vertex per clock to keep up. This means a maximum of 48 Vec4+Scalar ops per vertex. A more well balanced scenario between vertex and pixel shading might mean 12 Vec4+Scalar ops per vertex. Not all that fancy, but also not trivial. The 500 million polygons per second number of the 360 looks a lot more reasonable than the 125 million number thrown around when the original XBox was announced, or the 67 million number for the PS2.

On the RSX side, running the same vertex shader...
12 ops (Vec4+Scalar) per shaded vertex gives us, 8 * 550 / 12 = 367 million polygons per second.

This is where the flexibility of CELL SPUs may come in handy.

VS unit ~ vec4 + scalar ~ 10 flops/cycle*0.55gHz ~ 5.5 GFlops

SPU ~ 8 Flops/cycle * 3.2 GHz ~ 25.6 GFlops

Depending on the CELL<=>RSX integration, using a couple of SPUs + VS units, you could push it closer to 1 Billion poly/sec I guess...and you'd still have those powerfull PS units for fragment shading. Furthermore, another SPU could be used for post-proccessing. So essentially the SPUs would load-balance with RSX...

And you can't tdo this on XB360 because?
Why do you think they let the GPU read and write from the CPU cache?

On both platforms one pof the easier ways (read most likely to be used) to exploit the prarllelism is to move part of the graphics workload to the CPU. I'm not convinced that doing it for all geometry is the right approach, but for some of the geometry it certainly makes sense.

 

[Pozer]

There are games on PS2 that have a lot more polys on cars than GT4. I forgot the numbers but I think resident devs can chime in on this. I think that one F1 game in particular was said to push around 18Mp/s by it's developers. Also, consider most of the Xbox racing games run at 30FPS (except for RSC/2) which effectively halves their poly/sec throughput. Frankly, I can't think of a game where you'd actually have 3x more polys/sec on Xbox than on a comparable PS2 game. More yes, but not nearly that much more.

I can't beleive this is till being argued. There's lot's of developers that say otherwise and peopel STILL don't listen. it's freaking hopeless.

Oh my, I love the argument that xbox has better graphics because it's games run at half the frames per second. Like it never occured to a ps2 dev to just cut the frames in half and corner the market on amazing looking ps2 software.

I fear when you wear your system like a badge you can rationalize anything.

 

[Titanio]

And you can't tdo this on XB360 because?

I don't think he was saying you couldn't. It's obvious MS is accomodating that if the developer so desires.

However, I do think it should be an easier investment to make on PS3.

 

[Qroach]

Pozer,

Oh my, I love the argument that xbox has better graphics because it's games run at half the frames per second. Like it never occured to a ps2 dev to just cut the frames in half and corner the market on amazing looking ps2 software.

Are you for real or do you just ignore every single bit of evidence any developer around here has EVER given you? you hear something from people that have developed on xbox AND PS2 and you simply CHOOSE to believe somehting else.

I fear when you wear your system like a badge you can rationalize anything.

I fear for fanbois that think they understand console hardware better then developers posting in the same thread. You can't seriously be so "un aware" to not see ERP, DeanoC and many other confirm the bloddy obvious!

 

[Qroach]

However, I do think it should be an easier investment to make on PS3.

...and what leads you to this conclusion? I think ERP knows far more about the PS3 hardware then many of us on this board. If your going to say something like this you should back it up with a little more than opinion.

 

[ERP]

And you can't tdo this on XB360 because?

I don't think he was saying you couldn't. It's obvious MS is accomodating that if the developer so desires.

However, I do think it should be an easier investment to make on PS3.

Not really it would probably be easier on 360, since you could just read the Verts from shared memory. On PS3 you're going to have to schedule the DMA transfers. Probably not a lot in it.

The hard part is going to be keeping the memory in check for transformed verts. 1 billion verts /60 is 16.7 million verts/frame at say 32 bytes a verts is >512Mb's of memory. So the only realistic way to do it is to push into a ring buffer and synchronise with the GPU, which is a lot more work that the actual transforms.

The other issue which might not matter is making sure you have some way to serialise submission and a deterministic submission order if you're using more than one CPU to process the verts.

 

[Titanio]

However, I do think it should be an easier investment to make on PS3.

...and what leads you to this conclusion? I think ERP knows far more about the PS3 hardware then many of us on this board. If your going to say something like this you should back it up with a little more than opinion.

"I do think" " should"

I say this because the SPEs seem like ideal vertex processors. And there are 7 of them - that kind of granularity might make it an easier investment, as I think dedicating one or two of those to such work might be a little more palletable than putting aside one core on X360.

It's not a conclusion, it's a thought in my head. I'm welcome to feedback on it and/or correction if I am mistaken!

Not really it would probably be easier on 360, since you could just read the Verts from shared memory. On PS3 you're going to have to schedule the DMA transfers. Probably not a lot in it.

I didn't mean easier as in easier for the developer, but in terms of divvying up power and/or execution units, Cell seems more granular, and the SPEs should be good vertex workhorses, no?

The other issues, I will have to bow to your commentary, and the the commentary of others (who know what they're talking about)

 

[Shifty Geezer]

And you can't tdo this on XB360 because?
Why do you think they let the GPU read and write from the CPU cache?

I didn't see Jaws anywhere saying this couldn't be done on the other platform. :?

Wondering though, could this be a sort of intentional 'compensation' for lack of unified shaders? If the GPU is vertex bound, press-gang a SPE or two to help out? That way the SPE's can still be used elsewhere but for a graphics whizzo game, turned to helping with the visuals. Perhaps this was an idea from the off?

 

[Pozer]

Pozer,

Oh my, I love the argument that xbox has better graphics because it's games run at half the frames per second. Like it never occured to a ps2 dev to just cut the frames in half and corner the market on amazing looking ps2 software.

Are you for real or do you just ignore every single bit of evidence any developer around here has EVER given you? you hear something from people that have developed on xbox AND PS2 and you simply CHOOSE to believe somehting else.

I fear when you wear your system like a badge you can rationalize anything.

I fear for fanbois that think they understand console hardware better then developers posting in the same thread. You can't seriously be so "un aware" to not see ERP, DeanoC and many other confirm the bloddy obvious!

I'm not sure you read what I wrote or got the sarcasm, regardless I'm not going to argue with you. I guess in your world the ps2 GPU is faster and more feature rich than the xbox GPU. That's cool. I love playing games on the ps2, GC, and Xbox so no skin off my back. My eyes tell me xbox games look better (on average) with some exceptions. Thats all.

 

[Megadrive1988]

without making a new thread just for a simple, general question, I'll just ask here: the last two console generations (PS1, N64) (DC, PS2, Cube, Xbox)
have mostly emphesised polygon performance, with Xbox taking the first step into shaders.

this coming generation, the emphesis is mostly on shaders.

is that likely to be the main emphesis for the generation after, or will something else become the main focus in graphics, other than shaders ?

 

[Squeak]

REYSlike rendering.

Hi lb.

 

[j^aws]

Let's see...

Assuming ~1 vert per triangle. 1 triangle set up per clock. The vertex shaders need to produce one vertex per clock to keep up. This means a maximum of 48 Vec4+Scalar ops per vertex. A more well balanced scenario between vertex and pixel shading might mean 12 Vec4+Scalar ops per vertex. Not all that fancy, but also not trivial. The 500 million polygons per second number of the 360 looks a lot more reasonable than the 125 million number thrown around when the original XBox was announced, or the 67 million number for the PS2.

On the RSX side, running the same vertex shader...
12 ops (Vec4+Scalar) per shaded vertex gives us, 8 * 550 / 12 = 367 million polygons per second.

This is where the flexibility of CELL SPUs may come in handy.

VS unit ~ vec4 + scalar ~ 10 flops/cycle*0.55gHz ~ 5.5 GFlops

SPU ~ 8 Flops/cycle * 3.2 GHz ~ 25.6 GFlops

Depending on the CELL<=>RSX integration, using a couple of SPUs + VS units, you could push it closer to 1 Billion poly/sec I guess...and you'd still have those powerfull PS units for fragment shading. Furthermore, another SPU could be used for post-proccessing. So essentially the SPUs would load-balance with RSX...

And you can't tdo this on XB360 because?
Why do you think they let the GPU read and write from the CPU cache?

I didn't say you couldn't!

There was a whole bunch of patens I posted from IBM last year that described this mechanism.

My response was to richardfeil who was describing the flexibility of unified shaders on Xenos and how you could approach 500 million polys/sec. Conversely I was describing what SPUs add to the pipeline and how you could approach the triangle set-up limit of 1.1 Billion polys/sec. Basically the SPUs have a load-balancing effect like Xenos but they are a link between the PPE and RSX.

On both platforms one pof the easier ways (read most likely to be used) to exploit the prarllelism is to move part of the graphics workload to the CPU. I'm not convinced that doing it for all geometry is the right approach, but for some of the geometry it certainly makes sense.

It's quite clear to me with XeCPU and Xenos. However with SPUs and RSX, I was expecting them to remove the VS units and only leave PS units in RSX. But if SPUs and VS units can work asymmetrically, then I suppose it's irrelevant and even provides more options...

 

[Iron Tiger]

And I heard that some XBox games go as high as 50M PPS textured and lit.

I don't think there's a game that came even close to that. The highest I've ever heard was about half of that figure.

Julian Eggebrecht of Factor 5 claimed up to 25mpps in Rogue Leader. In the 2 player coop version of Rogue Leader included in Rebel Strike, the action was essentially doubled. While the poly throughput may not have doubled, there had to be a significant increase. And I've never heard anyone say that the Gamecube was capable of pushing the most polys of the 3 current consoles.

without making a new thread just for a simple, general question, I'll just ask here: the last two console generations (PS1, N64) (DC, PS2, Cube, Xbox)
have mostly emphesised polygon performance, with Xbox taking the first step into shaders.

this coming generation, the emphesis is mostly on shaders.

is that likely to be the main emphesis for the generation after, or will something else become the main focus in graphics, other than shaders ?

The DC, GC, and PS2 are also capable of shader effects. The Xbox hardware has the most capability with its programmable pixel shaders, while the GC has fixed function shaders, and the PS2 and DC get by (or got by in the DC's case) on software tricks that run well enough due to the respective designs of the architectures.

And to answer your question, it's really up to the developers. I think it would be cool to see a stylized game that used millions of untextured polys per frame. Virtua Racing and Virtua Fighter would look so cool if resurrected in their old style, with a steroid injection. There's also much that can be done on screen space with pixel shaders and no vertices. 3D had been simulated in games for many years before polygons entered the fray. We could see some awesome 3D-esque worlds created in the next gen that don't use polygons at all. Or we can see some 2D games that have fidelity, vibrance, and imagination like nothing before.

 

[nickguy94]

"And you can't tdo this on XB360 because?
Why do you think they let the GPU read and write from the CPU cache?"

The big problem with that is, there's simply not enough cache on the CPU to hand off that much of a workload from the GPU to the CPU. The cache is already being strained by the 3 cores, you can't put that much more of a workload on it or else it's going to bottleneck what the 3 cores are doing, and thus slow them down.