Ah well that explains the numbers then, they must only be counting vertex shader ops. It still wasn't quite double but close enough given the margin of error in the slide.
Xbox One (Durango) Technical hardware investigation
- Thread starter Love_In_Rio
- Start date
- Status
Ah well that explains the numbers then, they must only be counting vertex shader ops. It still wasn't quite double but close enough given the margin of error in the slide.
True. They often refer to designs that they have since abandoned and/or future design options interpreted out of context. The best thing to keep an eye on I've found is the inventors. Tardiff, Grossman, Mejdrich, Tubbs, Stone, Kippman, etc are popping up in a lot of the patents that appear to clearly involve Durango and its associated tech.
This tiling patent outlines a method for hardware tiling that leans heavily on eSRAM and the display planes by the sound of it. The basic concept seems to be that they let the GPU process individual tiles several layers deep so it can do the processing and blending and compositing tile by tile instead of waiting for the entire layer of sparse tiles to be collected, reorganized, and then shipped off to the GPU for processing all at once. In the latter process, the GPU would seem to be idle while waiting for the layers to show up. In MS's method, it's busy processing tiles almost nonstop. Maybe others who know more about the tech can read it and improve my interpretation though.
I think their VR ambitions are more tied up in the Fortaleza glasses. Plenty of patents on those too. Illumiroom was cool, but they aren't going to realistically try marketing an Xbox projector when they are aiming to ship Fortaleza glasses for the living room in late 2014. It's just not a cool enough idea to bother bringing to market imho. It'd be a cool app to use on PC's connected to projectors though maybe.
Here's the patent for the Fortaleza glasses they plan on shipping in 2014, as per the Yukon leak: http://www.faqs.org/patents/app/20120105473
It's for both AR/VR according to Yukon. Could be huge if it turns out as neat as the patent suggest (eye tracking, Oculus Rift-style lightweight glasses + Kinect 2.0 = VR).
Nah, it's from late 2011. http://www.faqs.org/patents/app/20110304713
I think the first we heard about them was in some weird chat conversation someone on GAF posted about circa Jan 2013 where some guy claimed to have an IRC chat with a supposed dev. Shortly after portions of the chat were posted Arthur confirmed some of the info with his Durango Summit info, including clarifying the dev's comments on 'multiple hardware scalers'.
I'll add one thing t the danger of using patents to draw conclusions. Depending on which department your in at MS and the patent budget, you may well file patents on anything and everything even superficially related if you think there is any merit.
On of my friends must have 50 patent cubes (anyone from MS will get the reference) in his office, which he filed over perhaps 12 years, the vast majority of those were only superficially related to the work he was actually doing.
I haven't read the tiling patent and I don't intend to, but any mechanism for TBDR would rely on ESRAM and some sort of view composition, that doesn't mean the hardware in Durango is there for that purpose, in fact I'd expect a lot less ESRAM if that were the case.
The rough overview just given sounds a lot closer to the Talisman work that was done.
On of my friends must have 50 patent cubes (anyone from MS will get the reference) in his office, which he filed over perhaps 12 years, the vast majority of those were only superficially related to the work he was actually doing.
I haven't read the tiling patent and I don't intend to, but any mechanism for TBDR would rely on ESRAM and some sort of view composition, that doesn't mean the hardware in Durango is there for that purpose, in fact I'd expect a lot less ESRAM if that were the case.
The rough overview just given sounds a lot closer to the Talisman work that was done.
It's interesting that they used an object on the ground with viewers standing to show on the front page of the patent ... 2 meters is about the minimum distance I'd see stereoscopic AR glasses with non adaptive focal planes for the overlay work, this obviously does limit applications just a tad.
There's no reason you can't adjust the focal planes in these glasses. All that takes is applying a voltage across certain materials to adjust the index of refraction appropriately. The 2014 version of the glasses is intended for living room use as VR/AR. That is really exciting I think. The more general AR specific stuff is for 2015, the mobile version of the glasses.
ERP,
I'm not sure it'd be aimed at deferred rendering necessarily. Not sure what kinda hardware would aid in the rendering of deferred game engines. I don't think the Talisman stuff handled tile in the same way though. This patent reads like it's a totally new method for that. One where your GPU is processing individual tiles several layers deep instead of the GPU waiting around for an entire layer to be ready for processing all at once. The aim seems to be to keep the GPU processing at all times.
I think the patent for the display planes also mentioned something about performance gains related to processing small segments of the planes in localized patches. Maybe that is alluding to the same thing, where the patches are the tiles? Maybe I should paraphrase the patents as best I can and pretend some insider pal of mine leaked the info to me so you guys can feel free discussing the merits of my interpretations.
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.
Very true. I am sure there are patents that MS haven't fully realized. There are MS patents that targeted the 360 that MS has chosen to issue patent continuation because some of the novel implementations have yet to be put to use. And MS has one of the largest research arms in the world so it probably has issued hundreds of patents off that research that never goes anywhere.
However, I do it because it's very interesting and its a natural extension of a part of my employment which revolves around shifting through scientific journals to see if someone's work is applicable to the work I do. I find it a rewarding because it interesting to see what others are doing in areas different from my own. Digging through MS patents are just as rewarding because of my interest in tech in general and consoles specifically, while my work is biological in nature.
And while any one particularly patent may mean a lot or very little, you can get a good sense of what MS is at least trying to realize by broadly looking at their console based patents as well as their overall patents in general. Everything points strongly at Durango being a x86 design so a lot of non gaming patents may be applicable. MS's software and hardware patents relating to heterogenous designs have broad implications so I find myself reading through not only for insight on Durango but on insight of what is MS is planning in general.
Plus MS patents makes perusing through their research sites more manageable and that site's white papers, transcripts and slide presentations provides more context then the patents that are based on that research work. We are talking procedural textures/geometry, heliOS (a OS design for heterogenous cores) and a ton of other research that may or may not have Durango implications.
Going by astrograd's description above, my gut reaction is that MS explored TBDR and patented this to tie up the idea, but that doesn't necessarily factor into Durango's design decisions or hardware focus. I'm assuming devs will be free to implement whatever rendering engines they like. And that sounds like a software choice to me.
(looks at patent)
Looks like a software patent effectively patenting changing the order you render and composite images. What an amazing invention! With typographical errors making the explanation uncertain! (I hate the patent system. I hate looking at patents. Drives me mad!)
Okay, silliness of the patent office aside, I'm reading a wider catch-all patent. It's talking about low-power computing in mobiles and the like, which already use TBDR. And I also don't understand its value. The intention is to avoid poor memory utilisation, but straight 2D image composites are perfectly cache friendly. There's no reason to thrash caches and get memory stalls. Whether you render out all deferred buffers as full planes and composite, or whether you render and composite in tiles, memory utilisation should be pretty much the same (while actual rendering will become for less efficient).
I remain of the belief that the SRAM focus is cheap BW, not some special hardware approach to rendering. Tiling hardware is IMO more focussed on tiled textures and meshes (megatexture/megameshes).
(looks at patent)
Looks like a software patent effectively patenting changing the order you render and composite images. What an amazing invention! With typographical errors making the explanation uncertain! (I hate the patent system. I hate looking at patents. Drives me mad!)
Okay, silliness of the patent office aside, I'm reading a wider catch-all patent. It's talking about low-power computing in mobiles and the like, which already use TBDR. And I also don't understand its value. The intention is to avoid poor memory utilisation, but straight 2D image composites are perfectly cache friendly. There's no reason to thrash caches and get memory stalls. Whether you render out all deferred buffers as full planes and composite, or whether you render and composite in tiles, memory utilisation should be pretty much the same (while actual rendering will become for less efficient).
I remain of the belief that the SRAM focus is cheap BW, not some special hardware approach to rendering. Tiling hardware is IMO more focussed on tiled textures and meshes (megatexture/megameshes).
Even if you could it only helps a little ... you still get the problem that if you have a lot of depth in the virtual image, then at close range not everything should be at the same focal plane. In VR this doesn't matter much, we can get used to a mismatch ... but in AR you will get the situation that a real world part of what you see and a virtual part should both be in focus, but won't be.
You could use gaze tracking to correct for this ... but you're fast getting to the bleeding edge of what's possible in the near future.
PS. adaptive lenses are not generally made with electro-optic effect materials.
SenjutsuSage
Newcomer
It's very possible that you're correct, but an inevitable consequence of their decision is very likely to be that developers find interesting ways to take advantage of potential benefits of SRAM, don't you think?
The patent covers this explicitly.
Adaptive lenses for typical eyewear, no. Screens intended to allow for projection onto them? Not so sure. Anyhow, I was only noting that patent as an example. It would be more interesting to discuss in another thread or closer to its supposed unveiling in 2014.
Shifty Geezer said:
Sure, but we know the DME's are designed around tiling to some extent and we have likewise heard that MS explicitly gave shout outs to Carmack and Lionhead for their implementations of virtual assets within the dev documentation about rendering on Durango. Plus MS has a long history with looking hard at tiling hardware. Add in the fact that we also hear murmurs of them consulting with devs on rendering trends (Arthur said as much on GAF iirc) and we know that devs on all the major game engines seem to be pushing towards virtualized asset rendering and it paints a picture that MS has likely implemented tiling for this kind of stuff at some level in their architecture. Will they leverage this new method for tiling to exploit this approach further? I am optimistic that they will because it makes a lot of sense with regard to murmurs we have heard in the past, long before this patent got published.
The patent explains why you don't need nearly as many memory accesses since you aren't copying data as often. And why do you say there's no reason to encounter cache thrashing in a 2D image? I find it hard to believe that the ppl inventing this at MS are making claims in the patent that are downright false. So when they say cache thrashing etc is avoided in advantageous ways as compared to other methods I'm inclined to believe as much.
bkilian suggested in the past that it wasn't just some patchwork addition intended to boost the bandwidth. It had other uses that are also important to the design. And who is saying tiling hardware isn't focused on those things? Processing the tiles in the GPU in more efficient ways is also important when the paradigm is highly inefficient. This wouldn't be the first time we've heard of the Durango GPU's processing being highly efficient and based on eSRAM latency. Cynics asserted that was just MS bragging about GCN in their documentation, but maybe this tiling method keeps the GPU processing almost continuously due to the low latency eSRAM and perhaps that is what they are referring to?
Just speculation, but it fits with everything I've read from insiders, the leaks, and the conclusions of devs saying the two are on par with each other while PS4 has a leg up in specs.
SenjutsuSage
Newcomer
Perhaps the old way of handling tiles isn't so much more prone to cache thrashing than anything else, which I don't think is the patent's intent, but in the instances where thrashing may take place with the old method, maybe the new method described in the patent would somehow significantly reduce such instances while potentially also being a more efficient way to handle tiling as well. But then, I too, am of the suspicion that there was more to Microsoft's thinking when they opted for ESRAM.
Last edited by a moderator:
Betanumerical
Veteran
The efficiency in the document you are referring to has to do with 1 Lane SIMD and 5 Lane SIMD. It was all about utilisation of these lanes and had nothing to do with the eSRAM or latency in any way.
SenjutsuSage
Newcomer
But couldn't the ESRAM, if it's 6T-SRAM, help out the GPU in some pretty important ways if developers utilize it properly? The ESRAM may not play a role in their claims for efficiency, but it could still turn out to be a whole lot more beneficial in real world games than some might be thinking at the moment.
I would like a justification for why it should matter at all to the developer if the 32MB pool is eDRAM or SRAM.
For all but a small number of conflict cases, the latencies of the two for a pool that size are generally equivalent, and in some possible implementations the DRAM is superior.
What technique is suddenly possible if the pool takes 28 cycles instead of 32?
SenjutsuSage
Newcomer
I can't provide that justification, as I'm not a developer, but previous posts here from individuals much more knowledgeable than myself seemed to imply that there was a chance that low latency 6T-SRAM might be pretty beneficial to Durango, or at the least made the architecture interesting.
And here is another post in this very thread.
http://beyond3d.com/showpost.php?p=1696970&postcount=245
Here is another over at GAF.
http://www.neogaf.com/forum/showpost.php?p=50467425&postcount=495
Quite a few more, but it's too tiring to search them all out. But you do get a general sense that depending on the type of ESRAM, that it could be a beneficial aspect of Durango's design.
I'm not sure but I think ERP suggested that there were two criteria that mattered. One was if it was low latency SRAM, 6T-SRAM, and the other was whether or not the ESRAM pool could be used as GPU input or something like that. I believe the ESRAM meets both those criteria, or at least I think it does.
Last edited by a moderator:
Strange, could've sworn others had claimed it was in relation to a specific shader utilization. Either way, on GAF Arthur had suggested that the processing itself was more efficient in some way, hence the flops comparison was inadequate. Based on what ERP said it sounds like the eSRAM could be leveraged for pretty significant efficiency gains.
SenjutsuSage
Newcomer
I believe it was said that a specific bit of code was run on both the 360 GPU and what I assume is Durango's GPU, or at least the same exact architecture that Durango's GPU will be based on, and the 360 ran it at 53% utilization and the hardware they are saying Durango will have, ran it at 100% utilization.
Betanumerical
Veteran
Thats exactly what it said but what it is referring to is a standard part of the GCN architecture.
I may be wrong but wouldn't a tiling based design be more latency sensitive than a conventional design. Trading latency insensitivity for the need for more bandwidth. The difference between 28 cycles and 32 cycles is 15%. Is that small enough to be insignificant?
SRAM also offers a more energy efficient memory, which may have influence the choice of memory. And doesn't the need for DRAM to be refreshed eat up cycles?
SRAM also offers a more energy efficient memory, which may have influence the choice of memory. And doesn't the need for DRAM to be refreshed eat up cycles?
Last edited by a moderator:
- Status
Similar threads
