More info about RSX from NVIDIA

nAo, their flop count also includes both the 4 way MAD + 2 flops SFU/scalar in the mini-fpu which they can't schedule simultaneously because the unit is only 4D; unless maybe they are counting bias/scale which is more non-programmable flops. So I'm only coming up with 16 programmable flops per PS ALU combo.

 

News.

Firstly, the Anandtech review is up.

Secondly, on the side this is a great read with a good overview of what the 7800 is bringing: http://www.1up.com/do/newsStory?cId=3141621

 

You're wish is my command: http://www.beyond3d.com/previews/nvidia/g70/

 

Just think of the relationship of an object and it's environment. You have three elements at play here: Vertex calculations, physics and pixel shading. I'm assuming that the RSX will handle the MAJORITY of vertex work, although I've read info that may suggest otherwise. However, for the sake of this discussion, I'll assume that the RSX handles the lion's share of vertex calcs.

How does a wave look when it crashes to the bow of a ship? Well that depends where it hits and the force in which it hits. In layman's terms, the Cell could handle the physics leg work and the RSX could use the results to accurately tranform vertices to correctly model the breaking wave. The fact that the RSX implements a form of TurboCache leads me to believe that the RSX will be an extension of Cell. They will work like one entity kind of like what Sony originally planned for with 3 Cell chips.

 

Has anyone confirmed that G70 = RSX with higher clockspeed?

 

RSX is not yet taped out , G70 is already on the shelf , i don't think rsx = g70

the WGF desktop composer and some DX9.0 legacy bound are not usefull in the rsx , plus i hope more pipe or alu like nv2a over nv20

 

Nvidia already said the archicture is similiar of rsx to g70, expect a 5% performance jump at best.btw its taping out in august , thats in 1 and a half month

 

So, if G70 = RSX, then total programmable flops per clock is:

RSX - 464 (no tex) or 272 (with 24 tex), max 8 vertex fetches & 16 pixels w/ 2xAA
Xenos - 480 (with 16 tex), max 16 vertex fetches & 8 pixels w/ 4xAA

Other differences: Xenos can still perform MSAA on floating-point frame buffers, and can still output double Z with MSAA enabled. RSX can perform FP16 filtering in texture units and not sure if Xenos can or requires pixels shaders to perform filtering. In Xenos all shaders have direct memory access, can use 5 components, and have filtered texture access.

 

Both supposedly have ~300M transistors. I agree that a lot of the PC-era stuff isn't necessary, though.

Perhaps one of the reasons it hasn't taped out yet is it's on 90nm on Sony's(?) fab plants, rather than 110nm TSMC?

 

reading the Baumann article the G70 cannot do Fp16 AA. its not a feature

 

Does that mean it can't do HDR AA?

 

how can you guesstimate 5% and not 4,8% or 6,1%

only the clock speed is 30% more than G70 , 550 over 420 so the performance will scale accordly.

plus the fixed resolution on 720p will let the coder to write shader width expectable behavior.

Let wait and see the first batch of game at TGS 2005.

 

What won't be needed in the RSX version of the G70 will be stripped out and transistors will be replaced with those needed to interface Cell and work with it.

 

Either you increase the clockspeed and decrease pipes or you keep the same clockspeed as G70 and increases pipes

 

Well, I do hope that nVidia does add the feature of MSAA with FP16 framebuffers. You'd think that'd be a required feature for all next-gen consoles.

 

Shouldn't the fixed resolution be 1080p on the PS3? All this talk about 1080p, surely all games will use it.

 

i still hope they dont use 1080p like default frame buffer , so precious MB frame buffer waste in that manner....