Watch Impress PS3 Technical article, from GDC (some new info)

[Carl B]

Thanks for that!

Sure no problem. I think a lot of people are confused about RAM speed vs RAM amounts, and it's not the first time I've heard the same question asked. In fact it's quite logical to view 200 MHz RAM and 800 MHz RAM of 1/4th the amount and think to yourself: "One-fourth the RAM and four times the rate: these are equal!" It's just that in reality there's more to it than that; there are times when the amount of data you can work with per second isnt what's important - because indeed for our example both RAM options are equal in that respect, theoretically - rather how much you can work with at any one given instant is what matters most, and the 64MB of RAM cannot work with more than 256MB of RAM can never work with more than 256MB worth of data in any one instant, no matter it's speed.

 

[Inane_Dork]

Certainly it's not as useful as in an open system, but it can be a relevant advantage, especially in the early days of a console when one still does not have a clear idea about the hw and how to take advantage of it.

I was under the impression that part of the appeal of Xenos is smoothing out cycle level hiccups in the pipe. I also thought that some uses of the GPU are naturally weighted, say shadow map creation, in such a way that they are accelerated with a unified approach.

Is that true or is it an oversimplification?

 

[dukmahsik]

so all we really know about RSX is that it can do 24 texture look-ups per clock, which I assume means it has 24 texture units.

we don't know for certain how many 'pixel shading pipes' or 'vertex shaders' it has, or how many ROPs it has.

so far, the best guesses are, 24 pixel shader pipes, and 8 ROPs.

what about this possibility: 12 pixel shader pipes, with 2 textures each, making 24 textures, 8 ROPs, and 6 vertex shaders. just a wild guess. probably wrong. but something has to give, given the 128-bit bit memory interface.

so if it does have 8 ROPS, why were we thinking it had 16 before? was 16 claimed?

what does that mean to fillrate then? halved?

 

[Titanio]

so if it does have 8 ROPS, why were we thinking it had 16 before? was 16 claimed?

what does that mean to fillrate then? halved?

Sony has announced nothing about ROPs or fillrate on RSX. It's effective fillrate, at least over to VRAM, would remain unchanged, as long as the ROPs can saturate the bandwidth (and 8 could).

 

[nAo]

I was under the impression that part of the appeal of Xenos is smoothing out cycle level hiccups in the pipe. I also thought that some uses of the GPU are naturally weighted, say shadow map creation, in such a way that they are accelerated with a unified approach.

Is that true or is it an oversimplification?

Not having worked on Xenos I can't really say whether what you say is true or not.
On the other hand a GPU employing an unified approach should provide all the good stuff you put in your list.
At this time I believe Xbox360 games will improve (gfx wise) at a lower pace than PS3 games (but I'm not saying anything about the starting point ).
PS3 will really shine when developers will start to use CELL to do clever things, not just to pump out more triangles.
In the end the old motto is still valid: the fastest triangles are those you don't draw.

 

[PSman]

I got a question, is the RSX capable of FSAA?.

 

[expletive]

At this time I believe Xbox360 games will improve (gfx wise) at a lower pace than PS3 games (but I'm not saying anything about the starting point ).

Interesting comment, could you elaborate on it? Are you saying that the GPUs will max out relatively early on in the lifecycles and the CPUs will be the differentiating factor down the road? If so, how would you chracterize any CPU disparity between the machines?

 

[Shifty Geezer]

Yes, but probably not in the way you are thinking

 

[expletive]

Yes, but probably not in the way you are thinking

Between you and Titanio with the teasing, cryptic posts i feel like this thread is a game within itself. Maybe i'm playing the "PS3 Spec ARG"?

 

[Mythos]

Sony has announced nothing about ROPs or fillrate on RSX. It's effective fillrate, at least over to VRAM, would remain unchanged, as long as the ROPs can saturate the bandwidth (and 8 could).

It has been annouced that RSX will..136 pixel-shader operations per cycle.

 

[Titanio]

It has been annouced that RSX will..136 pixel-shader operations per cycle.

That doesn't tell us anything about ROPs, though, right..?

BTW, I think that number was a mistake - nVidia quotes 136 shader instructions per cycle for the G70, that's probably what they meant to say for RSX also. Instructions and ops are different things, AFAIK.

 

[Brimstone]

Not having worked on Xenos I can't really say whether what you say is true or not.
On the other hand a GPU employing an unified approach should provide all the good stuff you put in your list.
At this time I believe Xbox360 games will improve (gfx wise) at a lower pace than PS3 games (but I'm not saying anything about the starting point ).
PS3 will really shine when developers will start to use CELL to do clever things, not just to pump out more triangles.
In the end the old motto is still valid: the fastest triangles are those you don't draw.

Why would 360 games improve GFX wise at a lower pace? I assume the Vertex processing capabilities of the CPU cores, the tesselator on the Xenos GPU, and MEMEXPORT will take some time to get familar with.

 

[Lycan]

I think nao meant that the improvement magrin for the the Xenos will be "wider"., which supports your assertion ! :smile:

 

[heliosphere]

I was under the impression that part of the appeal of Xenos is smoothing out cycle level hiccups in the pipe. I also thought that some uses of the GPU are naturally weighted, say shadow map creation, in such a way that they are accelerated with a unified approach.

Is that true or is it an oversimplification?

That's largely true. One area where Xenos has a big advantage is full-frame post-processing effects (HDR exposure functions, blooms, motion blur, etc.). For these effects you're drawing a single full screen quad (which means practically zero vertex shader cost) but running a pixel shader on every pixel (high pixel shader cost). On RSX the vertex shader is going to be sitting idle for most of the time it takes to run the post processing pass. On Xenos every available ALU is doing useful work.

There's some question over whether the overall benefit is really significant though - nVIDIA have suggested that by specialising the vertex and pixel shader units to their typical usage patterns they can get greater performance for a given number of transistors. It's hard to say for sure which will be more efficient for typical games. My feeling is Xenos will have the edge overall but we'll have to wait and see.

 

[heliosphere]

Is it true an entire SPE is dedicated to the OS when MS only asks for what 5% of two cores resources on Xenon?

Is that right? That OS sure seems hungry...hope it's doing a whole lot.

I don't think they reserve an entire SPE because they need all the power so much as it's impractical to share it for the kinds of things they'll want to use it for. I'm guessing they'll use the SPE for things like audio and video chat during games, both of which have fairly strict requirements in terms of latency / response time. It's very expensive to context switch an SPU and for these sorts of soft real time uses where you need to get some work done in a fairly short window to avoid glitches you can't really afford to share the SPE.

Being greatly simplified for the purposes of cramming as much floating point power as possible into a given number of transistors the SPEs don't support interrupts* or other hardware features to enable true pre-emptive multi tasking. A context switch has to be largely software controlled and requires swapping out the full 256K of local store (actually I'm simplifying slightly but that's true to a first approximation) plus backing up all 128 128 bit registers. It's very difficult to arrange sharing an SPE with arbitrary game code for anything that requires low latency access to resources or guaranteed response times.

* [edit] Actually there is some support for interrupt handling but it's not very sophisticated and the point still stands that context switches are expensive.

 

[Shifty Geezer]

Between you and Titanio with the teasing, cryptic posts i feel like this thread is a game within itself. Maybe i'm playing the "PS3 Spec ARG"?

I think you've confused me, as I replied to PSman before you had posted, but you replied before me, hence it looks like I was replying to you. It is to PSman asking about FSAA that I say yes, but not as he/she is thinking of it.

 

[ROG27]

That's largely true. One area where Xenos has a big advantage is full-frame post-processing effects (HDR exposure functions, blooms, motion blur, etc.). For these effects you're drawing a single full screen quad (which means practically zero vertex shader cost) but running a pixel shader on every pixel (high pixel shader cost). On RSX the vertex shader is going to be sitting idle for most of the time it takes to run the post processing pass. On Xenos every available ALU is doing useful work.

There's some question over whether the overall benefit is really significant though - nVIDIA have suggested that by specialising the vertex and pixel shader units to their typical usage patterns they can get greater performance for a given number of transistors. It's hard to say for sure which will be more efficient for typical games. My feeling is Xenos will have the edge overall but we'll have to wait and see.

It's hard to say. While Xenos's pool of ALUs may be more versatile...they are certainly weaker and, in total, fewer (48 for Xenos vs 56 for RSX). What will be more effective? ...RSX's independent shader pipelines which are harder to utilize efficiently, but will eventually yield predictable brute force in the closed box environment, or Xenos's efficient, load balancing unified shader pipelines that are lacking in brute force compared to the competition? As always, the realworld application will be the deciding factor...and non hypothetical numbers.

 

[blakjedi]

It's hard to say. While Xenos's pool of ALUs may be more versatile...they are certainly weaker and, in total, fewer (48 for Xenos vs 56 for RSX). What will be more effective? ...RSX's independent shader pipelines which are harder to utilize efficiently, but will eventually yield predictable brute force in the closed box environment, or Xenos's efficient, load balancing unified shader pipelines that are lacking in brute force compared to the competition? As always, the realworld application will be the deciding factor...and non hypothetical numbers.

What makes you say that the ALU's are weaker in Xenos than RSX? According to a certain ATI rep Xenos' ALUs are more powerful than R520s no comparison that I have seen has been done with R580/G70 however normalized for clock rate.

 

[dukmahsik]

It's hard to say. While Xenos's pool of ALUs may be more versatile...they are certainly weaker and, in total, fewer (48 for Xenos vs 56 for RSX). What will be more effective? ...RSX's independent shader pipelines which are harder to utilize efficiently, but will eventually yield predictable brute force in the closed box environment, or Xenos's efficient, load balancing unified shader pipelines that are lacking in brute force compared to the competition? As always, the realworld application will be the deciding factor...and non hypothetical numbers.

certainly weaker?

 

[!eVo!-X Ant UK]

That's largely true. One area where Xenos has a big advantage is full-frame post-processing effects (HDR exposure functions, blooms, motion blur, etc.)

All that can be done on PS3 aswel via SPE