Could be more RSX info...

"Multiway programmable shader processors"

(back home in Japan)

Whatever their finally be called I know I thing - RSX will have more than one processor core. Maybe not as many as Cell but it will have something.
 
leechan25 said:
Whatever their finally be called I know I thing - RSX will have more than one processor core. Maybe not as many as Cell but it will have something.
Care to enlighten us how you know this?
 
Xen said:
Notice he says a farm of these so called multiway programmable shaders with each able to process 136 shader ops a cycle simultaneously. I'll let you tech guys ponder the graphical possiblities. Certainly the 7800 GTX cannot do this.

G70 can do that.

Vertex Shader: 8VS*2-issue ALU = 16 shader ops
Pixel Shader : 24PS*2*2-issue ALU + 24PS*1 "free" nrm_pp = 96 shader ops + 24 shader ops = 120 shader ops

VS+PS = 16 + 120 = 136 shader ops.

btw, the RSX is "Multi-way programmable parallel FP shader pipelines", not "Multi-way programmable shader processors" according the E3 slides.
 
Last edited by a moderator:
Sethamin said:
There really isn't much information on RSX to go on, and in the absence of that information, I'm just assuming the simplest explanation is probably the correct one (see also: Occam's Razor...yes, I said it again!).

Try Hume's fork.....
 
Vennt said:
A better question would be why on earth you wouldn't want more?

Given the choice between a PS3 with a G70-like GPU or one with a G70.5++ like GPU I'd take the one with the most bangwhiz for my money.

But that's the exact problem - a bigger, better GPU would be a lot harder to manufacture, thus it would cost a lot more, and Sony wouldn't (and couldn't) swallow that...
 
RSX
~350 Million transistors
550Mhz
No vertex shaders (picked up by Cell)
No Purevideo
48 G70-class pixel shaders/TMU's
Intellisample
Ultrashadow

That would be stupid because the Vertex Shaders of the G70 are better for vertex processing than the SPE's.
 
Kutaragi made a conference in December where he talked about the future of Cell and 2 new variants of Cell, the Mini-Cell and de Micro-Cell. Mini-Cell is Cell with 4 SPE and Micro-Cell is Cell with 1 SPE.

I believe that the Visualizer will be the PSP 2 System Processor.
 
Sethamin said:
There really isn't much information on RSX to go on, and in the absence of that information, I'm just assuming the simplest explanation is probably the correct one (see also: Occam's Razor...yes, I said it again!).

Are you sure that a methaphysics argument in ontology against universals is a good argument for consoles...
 
Urian said:
I believe that the Visualizer will be the PSP 2 System Processor.

as much as I'd love that to happen, I think we still have a few years left until such a thing becomes feasable. I'm thinking maybe another 4 to 5 years and 32nm goes full production. Just my 2 cents though.
 
Are you sure that a methaphysics argument in ontology against universals is a good argument for consoles...
Yes, it seems pefectly acceptable to me. But fine, if you'd like it in plainer terms, when you assume, you make an ass out of u and me. Is that more amenable to you?
 
Guilty Bystander said:
That would be stupid because the Vertex Shaders of the G70 are better for vertex processing than the SPE's.

For vertex texture stuff maybe. SPE's should be great at regular vertex transforms.
 
the only thing we can guess at with some degree of confidence for PSP2 is that it will probably use some sort of low-power small Cell processor (mini Cell or micro Cell) and some Nvidia graphics technology. PSP2 will probably be as similar in design to PS3 as PSP is to PS2. the PSP2 should have alot of power - not as much as PS3 or X360, but probably more than Revolution. I agree, PSP2 is a long way off. 2010 at the soonest. we might see it before PS4 though :)
 
Are we all assuming each pipeline of RSX has a texture unit + dual shader units with mini-ALU's per shader unit? A lot of the estimated 10 million transistors per pipeline are a result of the different caches that the pipeline uses to help localize data, and help reduce external accesses. 24KB of texture cache per pipeline for example adds up to a lot of transistors. On a 24 pipeline G70, that's 576 KB of localized data. Nvidia raised the amount of texture cache on the G70 architecture when going from 16 to 24 pipelines, from 16 KB per pipeline to 24 KB. They may raise it again to 32 KB when going to 32 pipelines on their upcoming chips.

It possible that RSX will have some full pipelines, but also have some pipelines with texture units removed, and thus saving on a lot of transistors. I would not be surprised if RSX has 16 'regular' pipelines, and an additional 24 pixel shader pipelines, but still with the dual shader ALU's with their mini ALU's.. That would give RSX 80 pixel shader units, with 16 texture units, and 8 to 10 vertex units.

I'm just thinking out loud, and wonder what do you guys think? Can the TMU be decoupled from a pipeline to save on transistor cost, and thus offer more 'pixel shader' only pipelines?

The PS2's Reality Synthesizer 'GPU' offered this ideal, as I believe it had 8 'regular' pipelines, and 8 other simpler pipelines.
 
Sethamin said:
Yes, it seems pefectly acceptable to me. But fine, if you'd like it in plainer terms, when you assume, you make an ass out of u and me. Is that more amenable to you?

I agree with you that derivated part makes more sense , but this "one should not increase, beyond what is necessary, the number of entities required to explain anything" does not help none explaining that, in fact this recent quote is much better (also one will have a very hard time arguing that a realist is just making an assumption).

Anyway we can ask why would they do any big architeture diference in what already seems to be a great GPU? If someone is able to give a very reasonable and strong answer then it may be the case here there is a change of that but really I dont see none.
 
Edge said:
It possible that RSX will have some full pipelines, but also have some pipelines with texture units removed, and thus saving on a lot of transistors. I would not be surprised if RSX has 16 'regular' pipelines, and an additional 24 pixel shader pipelines, but still with the dual shader ALU's with their mini ALU's.. That would give RSX 80 pixel shader units, with 16 texture units, and 8 to 10 vertex units.

I'm just thinking out loud, and wonder what do you guys think? Can the TMU be decoupled from a pipeline to save on transistor cost, and thus offer more 'pixel shader' only pipelines?
Of course it "can", all you have to do is look at ATI parts - the texture units have been decoupled from the shader units in all their desktop parts since R300 and in Xenos as well (its more "pronounced" with Xenos and RV530+ though). However, for NVIDIA to do it it would take a fundamental shift in pipeline design and shader dispatch capabilities, which is unlikely to happen within a generation (especially given their texture pipeline can be seen very evolutionary throughout their entire history of parts).
 
Laa-Yosh said:
But that's the exact problem - a bigger, better GPU would be a lot harder to manufacture, thus it would cost a lot more, and Sony wouldn't (and couldn't) swallow that...

Sony broke the mold with the PS2 by resisting price drops and they were correct in doing so. I suspect the goal with the PS3 will be to create the perception that it is well worth the price even though it will retail for $499 (or $399) and will remain that way for quite some time even while Xbox 360 goes to a lower price point. The most successful consumer brands don't engage in price wars to win. Instead they re-inforce consumer perception that the product offers great value.

Obviously because Sony is competing with Microsoft they have to back up their rhetoric with a good product. The advantage Sony has is they already have mindshare from consumers, making it easier to out market the X-Box 360 if the PS3 has more potent hardware.

For Sony, marketing and hardware power have a symbiotic relationship when it comes to the PS3.
 
cho said:
G70 can do that.

Vertex Shader: 8VS*2-issue ALU = 16 shader ops
Pixel Shader : 24PS*2*2-issue ALU + 24PS*1 "free" nrm_pp = 96 shader ops + 24 shader ops = 120 shader ops

VS+PS = 16 + 120 = 136 shader ops.

btw, the RSX is "Multi-way programmable parallel FP shader pipelines", not "Multi-way programmable shader processors" according the E3 slides.


You know you just mistakingly made my point for me.

Each of these processors (and yes they were spoke of as processors not pipelines) can process 136 shader ops EACH per cycle simutaneously.

Do you not get it. Each of these processors equals the overall power (if clocked the same) as both vertex and pixel pipes of the G70

It amazes me that this information just gets glossed over so easily. Its like because the you cant compare it to any current cards out there (ie g70) this technology just cant exist.
 
Xen said:
You know you just mistakingly made my point for me.

Each of these processors (and yes they were spoke of as processors not pipelines) can process 136 shader ops EACH per cycle simutaneously.

Do you not get it. Each of these processors equals the overall power (if clocked the same) as both vertex and pixel pipes of the G70

It amazes me that this information just gets glossed over so easily. Its like because the you cant compare it to any current cards out there (ie g70) this technology just cant exist.

http://www.gzeasy.com/ours/edison/rsx_e3_slide_ken.png

http://www.gzeasy.com/ours/edison/rsx_e3_slide_huang.png
 
Back
Top