Xbox One (Durango) Technical hardware investigation

[Shifty Geezer]

It does look like 720 is slightly above the 1000 flops line, in line with 1.2 TF's rumored.

Only question is if Nvidia "knows" or is going on the same rumors everybody else is.

A quick google says Nvidia rates Titan at 4.5 teraflops. interestingly this is 2.5X the 1.8 TF PS4, not 3X *shrug*.

It's PR! I'd have thought people would understand what to make of PR by now (nothing!). nVidia will be no more informed than the rest of us, unless they have a mole working in AMD with complete access to all the hardwares. They clearly took the 1.2 and 1.8 TF figures and ran with them. They then compared their flagship model to the PS4 spec and, rather than bog people down with specifics, rounded the multiplier to make an more at-a-glance reference...oh look, it rounds up. Let's call it 3x bigger. We're not lying, just approximating using a viable communication convention.

PR is always PR. Don't trust those numbers. The only numbers you can trust are those from a hardware company reporting on their hardware, and explaining how they obtain those numbers.

 

[mosen]

And what is your source for this? MS's own patents are pretty explicitly suggesting otherwise and they explain a lot about the stuff we do know by the sound of it. Maybe the 'deferred' part is what's off? The patent below goes into detail about the procedure of rendering tile-based content and where this method's gains come from and it seems to mesh well with what ppl like Arthur on GAF had said months back. Additionally, the patent seems to directly allude to leaning on the eSRAM and display planes both for this methodology.

Here is the patent link for a new page...this link is a bit better actually since it has the diagrams:

http://www.faqs.org/patents/app/20130063473

I read through it and it sounds like there could be pretty considerable bandwidth and processing advantages rendering things on a tile depth basis as opposed to simply using tiles to construct layers, and then processing those layers in the GPU. This (new?) method could possibly explain why the eSRAM is targeted at such low latency and the murmurs from insiders of an exceptionally efficient GPU. It may not simply be a generic GCN setup making it "efficient" like many have asserted. It sounds like the more meaningful efficiency gains may come instead (or rather, in addition to...) the way the content layers are being processed.

Someone can correct me here, but I think in the typical approach to tile-based rendering support, even for stuff like the PRT support in AMD's recent hardware, the method for processing the content involved the GPU waiting around for the full layer to be stored in memory before it processes it all at once. This leads to the GPU sitting there with nothing to do in this capacity in the meantime, no? As such your GPU efficiency is bound by latency somewhat.

In the manner employed by the patent there, instead of your GPU waiting until an entire layer full of tiles is ready it handles the processing on a per tile basis as those tiles are stored in memory. The result is your GPU processing is bound by the latency of the eSRAM, which reportedly is extremely low (which is a good thing). At least, that's what it sounds like to me.


The image planes seem to play a meaningful role in this process too, so I'll re-read that patent tomorrow maybe and see if it adds anything interesting.

please take a look at this links too:

Patent application title: [SIZE=+1]RAY TRACING ON GRAPHICS HARDWARE USING KD-TREES/2010[/SIZE]

Patent application title: [SIZE=+1]MULTI-LAYER IMAGE COMPOSITION WITH INTERMEDIATE BLENDING RESOLUTIONS/2010[/SIZE]

Patent application title: [SIZE=+1]UPDATING GRAPHICAL DISPLAY CONTENT/2011[/SIZE]

also I find that one of those images that used in first link is kinda similar to this image from more recent patent, but maybe this similarity is not so important. :|

Patent application title: [SIZE=+1]SYSTEM AND METHOD FOR LAYERING USING TILE-BASED RENDERERS/2013[/SIZE]

http://www.faqs.org/patents/img/20130063473_06.png

 

[astrograd]

Who can know for sure, but I tend to agree that this makes the most sense. At least I think betting on this possibly being at least a fairly important part of their design calculations is a much safer bet than a surprise enhancement of the existing rumored specs. And plus it sounds damn interesting. I would really like to see a situation where they truly did have enough confidence in this idea of theirs to actually take it forward.

They went pretty hard back in the day with Talisman so maybe this is their more refined approach to that concept. I wish more of the tech folk here would bother reading these patents because it is very obvious they are referring to Durango in lots of them. I think ppl wanna just ignore them since they are so tedious to read. I also want to see ppl pay more attention to what Lionhead's Milo stuff has going on tech-wise. Their MegaMeshes stuff sounds like virtualized geometry as well as textures, which is what Carmack has been working on for Doom 4 iirc. I think that tech needs more attentino because the rumors seem to suggest Durango may be leveraging Milo-esque cloud AI for handling search queries in the cloud (aka the "Siri-like" voice interaction we've heard about). Would make the Kinect-required and always online stuff make a whole lot of sense imho. That and the MegaMesh/virtualized asset stuff from the same project makes me think Lionhead was actually making a tech demo/experiment for the next gen Xbox, but running it in software on a 360. Maybe their tech convinced MS it was the future of graphics rendering and they went with that at the core of the architectural design.

I think ppl should reconsider the assumption PS4 will perform better in general. I think it may be plausible that when both are processing games that rely on virtual textures Durango could very well perform significantly better. If that rumor about MS taking lots of dev input and designing their architecture around where the rendering trends were headed was legit then such games may be rather commonplace.

Anyone else here reading the patent? Does anyone know much about how hardware handles tiles/layers in the typical AMD/PRT setup today? Do we know if that setup handles the whole layer first or does it process individual tiles across all the layers? I seem to recall in the patent for the display planes (or the VGLeaks article, can't recall) that there is a performance boost when the content is placed in the image planes and processed in a highly localized fashion. The way this new patent reads sounds like it would also provide for that kinda thing.

Here is the basic rundown of the old/new procedure in a diagram:

http://www.faqs.org/patents/imgfull/20130063473_05

Old procedure is on the left, new is on the right.

 

[SenjutsuSage]

I'm trying my damndest to read through the patent, and there are times when I really think I'm taking it all in and understanding where it's all headed, and then a few lines appear that put me in my place, casting into doubt everything I just finished reading, forcing me to reread what I already read.

 

[BRiT]

Some of us tech folks are prohibited from reading patents, as it would put our respective companies in harms way in any possible patent lawsuits. It goes towards willingly and knowingly violating patents. It makes it nearly impossible to prove we discovered something on our own when we read other companies patents.

 

[astrograd]

Yeah it does a lot of repeating itself (it has to, unfortunately). The gist is (from the best I can gather) that it seems to handle tiled rendering in a new way that is significantly more efficient in terms of bandwidth consumption and requires less computations.

Additionally, (and arguably more importantly) it does it in a way where the GPU isn't sitting idle waiting for a content layer to fully (or sparsely in the case of virtual textures/geometry or SVO applications) written to memory before processing that layer. It also doesn't need to wait for several layers to accumulate before it can start blending them together for the final rendering. So less waiting around before actual processing can begin. This would presumably make it dramatically more efficient in handling tile based rendering approaches than traditional GPU hardware (including the PS4 and AMD's previous designs including PRT support). I think. Could be wrong though.

 

[RudeCurve]

I think it would be very cool if Durango uses some Talisman technology.

 

[astrograd]

Some of us tech folks are prohibited from reading patents, as it would put our respective companies in harms way in any possible patent lawsuits. It goes towards willingly and knowingly violating patents. It makes it nearly impossible to prove we discovered something on our own when we read other companies patents.

No kidding? Huh. Didn't know that. What about good ole 'random forum posts'? Let's say I posted my laymen interpretation of the patent and you, playing the role of moderator, read my interpretation. Would you likewise be bound to not read my post or reply to it?

 

[astrograd]

I think it would be very cool if Durango uses some Talisman technology.

I'm not so sure there is a whole lot left of that tech that hasn't already made its way into various GPU designs in some fashion. Not sure if any GPU does all of those concepts in practice or not though.

 

[SenjutsuSage]

Some of us tech folks are prohibited from reading patents, as it would put our respective companies in harms way in any possible patent lawsuits. It goes towards willingly and knowingly violating patents. It makes it nearly impossible to prove we discovered something on our own when we read other companies patents.

Oh yes, that's true. I remember reading about this. In that case, worry not, we're on the job. When some of us noobs get done reading these patents, there will be no puzzle left unsolved, no detail not taken completely out of context. And when all is said and done, we will extrapolate from the information within a console so powerful that it will blot out a thousand suns!

 

[dobwal]

Some of us tech folks are prohibited from reading patents, as it would put our respective companies in harms way in any possible patent lawsuits. It goes towards willingly and knowingly violating patents. It makes it nearly impossible to prove we discovered something on our own when we read other companies patents.

Yeah, thats a pretty reasonable justification for avoiding patents. That being said there a ton of microsoft patents revolving around tiling technology. But those patents by themselves don't mean that MS is employing those tech patents on Durango. MS has plenty of reason of exploring this type of technology outside of console gaming. This tech is applicable to their phones, embedded systems and tabs of which they more willing to take on a more hands on approach in the design and manufacturing.

It should be said that a majority of these patents don't explicitly list a gaming console as the target tech. And mobile devices are readily described in the patents themselves. However, they are a some that describe tiling tech with a game console listed as possible embodiment of the described tech and there is at least one that explicitly describe tiling on a gaming console.

GPU TEXTURE TILE DETAIL CONTROL
http://www.google.com/patents/US20120038657?printsec=drawing#v=onepage&q&f=false

This patent predates the QOS console patent from 2012 that showed a few console configurations with multiple gpus and cpus on board. However, it is actually the closest patent image I have found that somewhat resembles Durango (Fig 8). It eschews the multiple application and platforms cpu/gpu designations seen in the QOS patent. Its sports a few differences in that system memory is accessed throught the gpu and the gpu specific memory and controller aren't embedded or don't seem to be. But the basic components seem to be the same with the layout a little different. It might hint that when MS dropped the QOS design it fell back to an older overall design while incorporating some of the newer features of the QOS design.

 

[astrograd]

dobwal, the other patent we have discussed also explicitly links itself to being for a game console first and foremost. That one you linked to does go into detail about the operation of such a games console though.

There are loads of Kinect 2.0-realted patents that also do as much pretty explicitly. Oh, and that exact image (Fig. 8) has popped up in several of the patents I've been reading.



Just noticed this comment by Cevat Yerli (Crytek's CEO) in another thread here. I'd seen it before but didn't pay much attention to it (the guy is obnoxious as all hell when it comes to his PR speak). Anyhow, his comment here is interesting in hind sight. He is talking about next gen console tech and says the following:

Yerli opined that "this kind of realism will be playing a big focus in the next generation. And higher resolution texturing, like in the film industry, where you have effectively no limits on texture sizes.

"In the games, you used to have 256 by 256 textures, those limits will go up to...pretty much as much memory as you have in your PC. The new texturing will become a big topic as well."

Sounds like Crytek also feels virtual texturing is going to be the big trend for next gen. MS might be onto something...? Would make sense for them to get input from Lionhead and Crytek pushing for a more thorough hardware support for virtual textures/assets. We know how Carmack/id feel about it. How about DICE/EPIC? Anyone know how they feel?

 

[dobwal]

dobwal, the other patent we have discussed also explicitly links itself to being for a game console first and foremost. That one you linked to does go into detail about the operation of such a games console though.

There are loads of Kinect 2.0-realted patents that also do as much pretty explicitly. Oh, and that exact image (Fig. 8) has popped up in several of the patents I've been reading.

I am just highlighting my experience. I've read through dozens upon dozens of MS patents some applicable to gaming consoles some not. And the one thing I learned is that there are dozens of MS patents that possibly links to Durango because of specific tech descriptions but don't explicitly list a gaming console as embodiment for the patents. But hints may exist just as game pad, camera or sensor that allude its may be used for more than a generic PC.

MS basically states that their patents don't describe all the embodiments suited for the novel technology that are their patents describe so my exercise of patent reading is very wide.

 

[Shifty Geezer]

MS basically states that their patents don't describe all the embodiments suited for the novel technology that are their patents describe...

That's true of all competently drawn up patents (and not true for many amateur inventor patents who fail to secure what the inventor considered obvious interpretations). A patent in a company is about securing an 'idea', which may have been developed while exploring ideas for a particular product but which may go on to be used elsewhere, or nowhere.

I'm curious what percentage of patents investigated over the past decade relating to consoles have ever actually amounted to anything. I can't really remember any. I think there was a Kinect one. Move and Wii. The Cell ones, but then we had far better information on that than needing to dig through patents. We've seen plenty about techniques like 3D scene maps that some were adamant would feature in Wii and splitable controllers and fancy processors (Cell visualiser) that were total non-events. Just throwing that out there as a reference for people when looking at patents.

 

[astrograd]

I'm curious what percentage of patents investigated over the past decade relating to consoles have ever actually amounted to anything. I can't really remember any. I think there was a Kinect one. Move and Wii. The Cell ones, but then we had far better information on that than needing to dig through patents. We've seen plenty about techniques like 3D scene maps that some were adamant would feature in Wii and splitable controllers and fancy processors (Cell visualiser) that were total non-events. Just throwing that out there as a reference for people when looking at patents.

It's not merely random patents that may be related to a general computing environment we are talking about. The stuff I'm talking about is all stuff that directly references the particular architecture rumored for Durango and specific features of the platform that have also been rumored/leaked. There's several Fortaleza patents out there. Several for a variety of Kinect-related inventions. Patents for the display planes and the general platform architecture itself as well. Yes, there are TONS of patents out there but some are clearly intended to cover Durango. Unless the rumors are all bogus, which is unlikely.

I do agree ppl should be careful to look for specific mentions of things unique to Durango when patent snooping though.

 

[SenjutsuSage]

That's true of all competently drawn up patents (and not true for many amateur inventor patents who fail to secure what the inventor considered obvious interpretations). A patent in a company is about securing an 'idea', which may have been developed while exploring ideas for a particular product but which may go on to be used elsewhere, or nowhere.

I'm curious what percentage of patents investigated over the past decade relating to consoles have ever actually amounted to anything. I can't really remember any. I think there was a Kinect one. Move and Wii. The Cell ones, but then we had far better information on that than needing to dig through patents. We've seen plenty about techniques like 3D scene maps that some were adamant would feature in Wii and splitable controllers and fancy processors (Cell visualiser) that were total non-events. Just throwing that out there as a reference for people when looking at patents.

Well, don't forget the display planes patent, which is in Durango. Best example I could think of. It also has the exact same number as what's inside Durango described in the patent, too.

 

[Shifty Geezer]

It's not merely random patents that may be related to a general computing environment we are talking about. The stuff I'm talking about is all stuff that directly references the particular architecture rumored for Durango...

I haven't read the patent and it may be relevant, but I just wanted to throw that out there because we've been seeing a lot of patents regards next-gen hardware, and some are clearly not related but people still interpret them as important. It's like patents are seen as a gateway into the closed doors of the hardware design process. Which is true if you find the right ones, but "here's a patent for Illumiroom - Durango virtual reality confirmed!!1!" is jumping the gun, and a lot of people end up jumping the gun!

Did the display planes patent come before or after the rumour of their inclusion?

 

[SenjutsuSage]

I haven't read the patent and it may be relevant, but I just wanted to throw that out there because we've been seeing a lot of patents regards next-gen hardware, and some are clearly not related but people still interpret them as important. It's like patents are seen as a gateway into the closed doors of the hardware design process. Which is true if you find the right ones, but "here's a patent for Illumiroom - Durango virtual reality confirmed!!1!" is jumping the gun, and a lot of people end up jumping the gun!

Did the display planes patent come before or after the rumour of their inclusion?

After the initial appearance on a vgleaks poll, but before the actual vgleaks article detailing their purpose.

 

[pjbliverpool]

The full slide offers a bit more perspective, as the zoomed one makes the PS4 look 2X as powerful

It does look like 720 is slightly above the 1000 flops line, in line with 1.2 TF's rumored.

Only question is if Nvidia "knows" or is going on the same rumors everybody else is.


A quick google says Nvidia rates Titan at 4.5 teraflops. interestingly this is 2.5X the 1.8 TF PS4, not 3X *shrug*.

1.84X3=5.52

It's interesting they correctly give credit for Xenos perhaps being slightly more powerful than the best PC GPU's at it's release. At least it was in the same small ballpark. A notable accomplishment we aren't likely to see again.

Xbox was 2X geforce 3? Not sure on that one.

NV are definately taking a few liberties with the numbers there. I'd assume the 7800 comparison is referring to the 7800 GTX at 200 GFLOPS (less than PS360) rather than the more powerful 7800GTX at 255 GLFOPS (marginally more than PS360). But the NV2a comparison is miles off. In actually it had roughly 37% more FLOPs than the GF3 - and both were over 10!

 

[Gubbi]

But the NV2a comparison is miles off. In actually it had roughly 37% more FLOPs than the GF3 - and both were over 10!

Also, FLOPS is a poor measure of performance in the DX8 era with pixelshaders being integer based.

I'd also like to see the performance normalized with power consumption. PCs are transforming into CE devices, where the compact form factor dictate the power envelope to be less than 100W.

Is it relevant to compare peak GPU PC performance when less than 5% of all PCs have that kind of performance?

They should rename it "Console vs enthusiast gamer PC/workstation over time"

Cheers