RSX = Stream Processor!?!

Mintmaster said:
$20 says those "3+ years" includes development of the NV40/G70 architecture. borntosoul has it down perfectly

Hardly worth it, thats like £9-10 ;)

All jokin aside, you maybe right, but why would Nvidia AND sony make the G70?? Would'nt it of been cheaper if Sony let Nvidia make G70 on there own??? That way it would cost Sony alot less as they only have to modify the G70 for PS3??
 
Memory bandwidth is a concern, no doubt. I'd love one of the improvements/changes to RSX from E3 to be a move to XDR from GDDR-3.
 
Mintmaster said:
Well, a 550MHz G70 is 2.1 times the speed of a 400MHz NV40 in pixel shading and texturing, so voila! The only problem is that it has just over half the bandwidth of one of those NV40's.

To local memory yes, but it Also has the same again (and then some) going to system memory. All your static geometry and textures can go in local vram, whereas anything dynamic (procedural geometry, high-def video etc) can stay in system memory, and render straight from there.

You can also balance the load between the two. If you hit bandwidth limits to local memory, then move some of the textures into system memory and render from there. A multi-texture op could have 3 texture channels in local memory and 3 in system memory, sharing the bandwidth between the two paths.
 
Barbarian said:
More like post-January but yes indeed. What was disclosed at E3 is still true but obviously not all inclusive.
And they did improve on a lot of things, but like I said in an evolutionary way. I'm mostly excited about the software side of things, since Sony has decided to provide a down-to-the-metal API which makes me very happy.

Well this thread sure took a turn for the better. Thanks so much for sharing this nugget with us Barbarian ;)

Now let the speculation begin. From his comments he states "what was disclosed at E3 is still true", so I'm guessing the shader ops, clock speed and other things are the same from E3, but aspects that we don't know about were improved compared to a G70. Any guesses from anyone? I see xbdestroya thinks the ram speed will be upgraded. Anybody else?
 
Last edited by a moderator:
Wouldn't 32pipes throw off the DOT products count from E3? I remember a fun little thread last year where we were speculating on the VS:pS config based on DOTS and shops. It would seem that according to Jaws' conclusions and Goto's calculations that 8:24 worked best. What about a 24-pipe G71? Tranny count stays near that 300M mark, and they probably get the redundancy for better yields. Uh...or something. ;) Time to check back in on that G71 thread on the PC forum. PEACE.

EDIT: This thread: http://www.beyond3d.com/forum/showthread.php?t=24211&highlight=product
 
Last edited by a moderator:
MechanizedDeath said:
Wouldn't 32pipes throw off the DOT products count from E3?
At 32 pipes there'd be some redundancy in there for sure, which'll be a max of 28 pipes useable.

Regard Barbarians comment
And they did improve on a lot of things, but like I said in an evolutionary way
The term Evolutionary is fairly ambiguous. The theory of evolution covers both new species from old and small variations between generations. Is RSX a new species derived from G70, or a next generation with a couple of minor variations? I presume Barbarian means the latter, which I'd take as things like a clockspeed increase or a couple extra useful functions in the shaders. But then again evolutionary could mean taking the same PS and VS and restructuring them into an entirely new beast. Once again, no solid info! ;)
 
MechanizedDeath said:
Wouldn't 32pipes throw off the DOT products count from E3? I remember a fun little thread last year where we were speculating on the VS:pS config based on DOTS and shops. It would seem that according to Jaws' conclusions and Goto's calculations that 8:24 worked best. What about a 24-pipe G71? Tranny count stays near that 300M mark, and they probably get the redundancy for better yields. Uh...or something. ;) Time to check back in on that G71 thread on the PC forum. PEACE.

EDIT: This thread: http://www.beyond3d.com/forum/showthread.php?t=24211&highlight=product

Nah, 8:24 was inconsistent with shader ops/cycle and Dots /cycle. I actually voted for option C in that poll. I.e. 4 VS and 24 PS, which is consistent only if the VS units are upgraded to being 4 issue instead of 2 issue.

This would fall inline with the evolutionary comments. It would also make each VS unit more capable and closer to PS units. It would also follow that both VS units and PS units are evolving to become more similar and pave the path towards future 'unified' units in perhaps the G80...
 
Hardknock said:
I see xbdestroya thinks the ram speed will be upgraded. Anybody else?

Hey hey, not I think, rather I hope! ;)

And the move to XDR isn't for speed reasons per se, but because you can have more modules/devices pin for pin, greatly increasing potential bandwidth.
 
Jaws,
Why only 4 VS Pipes?To offload them for the Cell's SPEs?

What some of you said about a cut down version of the G71 made sense.Now coming back to the argument, if it was just a G70 again I repeat wouldn't Sony had announced this much earlier rather than to wait until somewhere near the launch of the G71?The G70 has been out for a while if development was moving on parellel course, shouldn't the RSX be out much earlier already?The most significant difference would only be the FlexIO and 90nm part right?Did they really need that amount of time?What advantages does Sony have if they kept quiet for all these while?Having an inferior RSX would mean better marketing for the G71 instead.

Disclosing the RSX earlier could have probably given the public a clearer idea of how Nvidia's next part will look like.Why so?The G71 is already a full 32pipe solution while the RSX may just be a customized G70.Could it influence the launch of the G71?
 
Nurbs?

Could this "Evolutionary" step be more towards a *form* of doing Nurbs rendering?

Modifying a rasterized surface, such as by trimming: File date March 2, 2004 & Issue date December 15, 2005

Embodiments of methods, apparatuses, devices, and/or systems for modifying a rasterized surface, such as by trimming, for graphics and/or video processing, for example, are described.

2. The method of claim 1, wherein said surface comprises an NURB.

One issue that relates to graphics quality is the rendering of trimmed surfaces. In one approach, trimmed Non-uniform Rational B-spline (NURB) surfaces are rendered with Adaptive Forward Differencing. See "Rendering Trimmed NURBS with Adaptive Forward Differencing," by Shantz and Chang, Computer Graphics, Vol. 22, No. 4, August 1988, pp 189-198. In this approach, adaptive forward differencing is extended to higher order, the basis matrix for each scan is computed, the shading approximation function for rational surfaces is calculated, and the NURB surfaces are trimmed and image mapped. Trimming is accomplished by using AFD to scan convert the trimming curves in parameter space, producing the intersection points between the trim curves and an isoparametric curve along the surface. A winding rule is used to determine the regions bounded by the curve which are then rendered with AFD. In another approach, all trimmed surfaces are converted into individual Bezier patches with trimming regions defined by closed loops of Bezier or piecewise linear curves. Step sizes are calculated in parameter space for each curve and surface which guarantee the size of facets in screen space will not exceed a user specified tolerance. All points on the trimming curves where the tangents are parallel to the u or v axes are discovered, here, the local minima and maxima. Using the extremes, the trimming region of the patch is divided into u,v-monotone regions. Each region is defined by a closes loop of curves. Using the calculated step sizes, each u,v-monotone region is uniformly tessellated into a grid of rectangles connected by triangles to points evaluated along the curves. The polygons defined in u,v parameter space are transformed into facets in object space by evaluating their vertices with the surface factions. Surface normals are also calculated. Each facet is transformed to screen space, clipped, lighted, smooth shaded and z-buffered using 3D graphics hardware. See "Real-Time Rendering of Trimmed Surfaces," by Rockwood, Heaton, and Davis, Computer Graphics, Vol. 23, No. 3, July 1989, pp 107-116.

As illustrated by block 115 of FIG. 1, higher order surface tessellation occurs early in the geometry processing phase of a graphics pipeline. Higher-order surfaces use mathematical formulae and/or functions to represent three-dimensional (3D) surfaces. Examples include Non-uniform Rational B-splines (NURBs), Bezier curves, N-patches, and more. The data transferred is tessellated to generate more complex models. The GPU, therefore, dynamically generates or tessellates the primary model data from the application into much more detailed and complex geometry.

Likewise, a third embodiment or instantiation of dedicated graphics hardware shall be referred to here as a programmable streaming processor. A programmable streaming processor comprises a processor in which a data stream is applied to the processor and the processor executes similar computations or processing on the elements of the data stream. The system may execute, therefore, a program or kernel by applying it to the elements of the stream and by providing the processing results in an output stream. In this context, likewise, a programmable streaming processor which focuses primarily on processing streams of fragments comprises a programmable streaming fragment processor. In such a processor, a complete instruction set and larger data types may be provided. It is noted, however, that even in a streaming processor, loops and conditional branching are typically not capable of being executed without intervention originating external to the dedicated graphics hardware, such as from a CPU, for example. Again, an embodiment of a GPU with this level of capability or a similar level comprises a programmable streaming processor in this context.
 
hugo said:
Jaws,
Why only 4 VS Pipes?To offload them for the Cell's SPEs?
...

This is my speculation based on E3 numbers being correct for RSX and CELL, i.e. 136 Instructions/cycle for RSX and 52 DOT products/cycle for RSX (deduced in the links above). These 4 VS units, each being 4 issue instead of 2 issue (G70 VS) and 24 PS units, still 5 issue, would be consistent.

Yes, I'm thinking along the lines that CELL has vertex power to exploit along with these, fewer, but more capable 4 VS units. A possibility being that they may be adapted/ evolved for geometry shading to work alongside with SPUs... though that may fall under being "exotic"!

Here's a thread on an NV patent for a programmable geometry shader,

http://www.beyond3d.com/forum/showthread.php?t=20445

EDIT:

Patent link fixed,

User programmable geometry engine
 
Last edited by a moderator:
Tahir2 said:
Yep, can't get away from that problem of memory bandwidth..

Edit: same applies to all consoles!
enough complaints can do wonders ;)

All I hope is that the RSX and whatever changes where made allow it to at least keep up a bit with the latest h/w this year, that is it's performance shouldn't be worse than a much older gpu on an older process, this'd be the first time that'd happen with sce h/w.
 
Maybe it has parts and features of Both G70 and G71??

You could have two groups of engineer's working in a lab, the first set of people are making the G70 and G71, and the other team are making RSX.

Team 1 = G70, G71
Team 2 = RSX

Now as Team 1 creates the new technology for its project and implements it, team 2 could come over have a brouse and see if it would work with there design goal's. That could be the reason why RSX is'nt done yet, Because G71 is'nt ready ;)

Also this way would tie into the whole tech "Sharing" that RSX and G70/G71 are suposed to have.
 
Jaws said:
So that sounds like a G7x and not G70, *shrugs*.

It may well be it's own "G7x", G7PS3 or something :p

Seriously, though, my expectation is they took G70 tech (a 8:24 setup), shrunk to 90nm and tweaked it quite a lot a lower levels to tailor to Sony's wants/needs (in a similar way to a G71 being a shrink with perhaps lower level tweaks).
 
Jaws said:
An old NV, Tony Tamasi Geforce 7800 video Interview, around 5 minutes in,

"...so just like Geeforce 6, we had a variety of products, well, we have the 7800 GTX and then the next, future generation of this technology will be the RSX..."

http://media.gear.ign.com/articles/628/628947/vids_1.html

So that sounds like a G7x and not G70, *shrugs*.

I cant view the video, because i aint registered there :( is there away to get it with out registering ?
 
Titanio said:
It may well be it's own "G7x", G7PS3 or something :p

Seriously, though, my expectation is they took G70 tech (a 8:24 setup), shrunk to 90nm and tweaked it quite a lot a lower levels to tailor to Sony's wants/needs (in a similar way to a G71 being a shrink with perhaps lower level tweaks).

I know what you mean but reading what he said,

"...so just like Geeforce 6, we had a variety of products, well, we have the 7800 GTX and then the next, future generation of this technology will be the RSX, so kinda what, what we characterise as the parent technology is this second generation shader model 3 engine. And it can be extentiated in a variety of ways, and in the case of the RSX, we have a many pipeline chip which would be directly coupled to the CELL processor..."

It sounds like a future tech, G7x -> modified into RSX, rather than G70 -> modified into RSX. And this would naturally be evolutionary...

EDIT:

!eVo!-X Ant UK, you can download the low res versions without registering... well I did ages ago...
 
Last edited by a moderator:
Back
Top