News & Rumours: Playstation 4/ Orbis *spin*

Timothy Lottes is an nVidia employee with no insider knowledge of either platform. That blog is his personal speculation based on his interpretation of rumours, backed with considerable expert understanding, but also with the same big question marks over the architectures that we're all shouldering. It is nothing like definitive, and shouldn't be treated any differently to any technical discussion we have here with developers. It's quite possible he's very wrong on some points.

 

Maybe he really know what he is talking about?

http://www.ign.com/articles/2013/01...dly-leaked-to-nvidia-by-former-amd-executives

 

Well if that's true, his opinions of Durango may well be true (VLIW instead of GCN), which would make a substantial difference to Durango. That goes against other news we're hearing though. Impossible to see, the future is.

 

Why throw around baseless speculation as if he has some magic insight? He does not. Please stop.

 

I assume Timothy doesn't, but if nVidia have had details leaked to them (which this AMD lawsuit claims), then perhaps he has?

 

Well ign is still clinging to the 6670 rumors, If that's what they got from nVidia I wouldn't be worried.

 

It's possible but highly unlikely. As anyone with potential knowledge of anything on those papers that came over from AMD should be smart enough to know that they can't do anything that might imply that they have knowledge of anything contained in those papers.

Hence anyone saying anything based on those papers in a public forum (such as an interview or website) would have to be colossally stupid, especially with the pending lawsuit.

So, it's far more likely he's speculating just like the other developers on these forums that don't have access to a durango dev kit. And even if they had access, they probably wouldn't be able to talk about it other than in very general terms.

Regards,
SB

 

When people leave a company and go to a competitor, they bring information with them. It happens. That's one of the reasons companies try so hard to poach people from their competitors.

That said, I don't necessarily believe that Lottes has any super-insider info.

 

And they're stilling clinging to the IBM PPC rumors for Durango.

 

A last minute switch to Kepler (or Maxwell) in Orbis? That would explain everything. *chuckles*

 

Yeah that would be extremely smart, providing additional material for the law suit.
Personally I think he's just speculating, there is nothing outlandish in his analysis, though he's very dismissive of Durango, and I'm not sure what he's basing the older architecture conjecture on.

It's easy to dismiss Durango as weaker if the leaks are true, less overall bandwidth, less CU's, if the intent of the ESRAM is to act primarily as a frame buffer as the embedded memory in 360 did IMO it is without a doubt weaker.
As I've said elsewhere if the DDR3 is intended to be the frame buffer, and the ESRAM the data source there is a chance it can offset some of the computational disadvantage, but it's still at at a bandwidth disadvantage and what I'm suggesting requires additional bandwidth to move data to the ESRAM before consuming it.
The question then becomes does it have enough bandwidth, something I can't answer because you need to actually use it to determine if it is.

 

Wow it has been a long time. I remember when we were doing this back in 05 and 06. Shifty Geezer is now a mod. Congrats ol friend. I didn't post that much but have much respect for you.

You guys mentioned stacking earlier. Vita has this tech right? So is it feasible we might see this in PS4?

 

Vita's stacking is a system in package setup that's not quite the same as the sort of stacked memory most have been talking about.
The system DRAM is physically on top of the SOC and VRAM, but it is wire-bonded.

The graphics RAM is Wide IO RAM whose pads attach to the face of the SOC below, basically by flipping the chip so that the logic layers for the SOC and memory are facing each other.
It's a minimal amount of stacking, and not the same sort of technique being discussed as 2.5 or 3D stacking.
It really only works for one extra layer, since no vias are being used to punch through silicon, and the chips being face to face means no additional chips can be joined in this manner.

The Vita's SOC is undoubtedly much lower power than whatever goes into Orbis, which makes me think this particular method won't work. It's no good to have an SOC that needs a full cooling solution insulated by memory attached in this fashion.

 

How would putting the framebuffer in main RAM be preferable? Framebuffer (Z-buffer, render targets etc) I/O is by far the biggest consumer of data when generating 3D imagery.

Vita's stacking is as primitive as you can get, standard wirebonding between the dies. It wouldn't work in a high-powered SoC because the interposer and DRAM on top of the main die would cause the package to cook, and you can't put the main die on top because you can't bring enough power and I/O to it through wirebonding; you need those BGA connections. However, through-silicon via tech is still immature and probably isn't ready - or at least cost-effective - for a system that would need to be manufactured in tens of millions of units per year.

 

I don't think the blog article has provided any additional detail. He was just talking about rumours. Similar to devs here talking about public rumours.

I now think he just hasn't read around so much. He's updated the blog:

His first mistake was to listen to SemiAccurate.

 

The bolded is what's important here.

 

Yes that used to be the case, and I can't see any reason you wouldn't leave Z in the ESRAM. Modern games read a lot of texels/pixel though, an I'm not sure that the FB is the primary consumer of bandwidth for say opaque polygons with a rough Z sort. 1080P with a deferred renderer with say 28 bytes per pixel requires about 60MB per full screen, if you say end up with 3x overdraw because of the sorting/ early Z it's 180MB, 60GB/s gives you 1GB/frame, so if it's BW limited ( and I'm not sure you wouldn't be ROP limited) it's about 20% for the opaque pass.

I'd also render the transparent pass into the ESRAM, but I'm speculating about a piece of hardware I will likely never use.
You'd probably also render shadows and possibly environment maps to ESRAM and then use them directly as a texture source.

Texture caches used to be almost perfect, but that was because everything was computed in the vertex shader and all the pixel shader did was interpolate, now a lot of the lighting computations are done in the pixel shader, and anything using a reflection vector to, look up a texture is likely to induce cache misses.

Having textures in the ESRAM is the only way I can see any offsetting of the rumored ALU deficit.

 

VGleaks: Orbis Unveiled!

 

Specs discussion over here please: http://forum.beyond3d.com/showthread.php?t=62936

There's the other non-spec stuff, of course. (Controllers, SKUs)

 

Move and evolved dualshock? It should be combined... but maybe it's like a Gobot and you can merge them together.