PS3: OpenGL ES V REYES V Raytracing/GI Pipelines + TBDR?

[j^aws]

He's from ArsTechnica and sounded like a developer from his other posts. If you bump into Simon, then send him this way! ...In his absence does anyone else know what he meant?

That the author didn´t have a clue obviously?

IMO, when Simon made that comment about Jason's post, it's kinda rude not to give any explanation to the uninitiated...or clarify the post. A simple 'he's clueless' is not constructive, though, I'm sure Simon didn't intend to.. ...So in his absence can you elaborate on what Simon was implying?

From the patent, it's 'tile' based alright. So it's atleast a TBR. So what's the definition of 'deferred' and 'non-deferred'? And what's a definition of anything in between? Hybrid, Semi?

Again accelerators have been using tile or chunk based memory optimisations for eons; fairly since the advent of 3D and yes even most if not all IMRs. A deferred renderer works ideally if all scene data gets collected prior to rendering and I think the former link provided illustrates it adequately.

Have you read the patent? Also please read my reply to Faf above. I'm not aware of any dealing with 'tiles' and overlapping 'bricks' to render scenes the way that patent describes.

Didn't want to bring any figures into this thread...no concrete evidence of the above, so pick a figure really!

But there´s "concrete" evidence of PS3 being a TBDR? I targeted and asked about antialiasing and sample densities on purpose, just because it´s one of the departments a DR has well known advantages. I recall one engineer here not long ago mentioning something about 64x sample SSAA with 64bpp HDR at 1024*768*32 for a real TBDR.

Hey, there are patents for those at least and I've already mentioned to prefix 'hybrid' to TBDR at the beginning of the thread and it's also the 'title' of another linked thread! ...If your really interested in bandwidths, these would be my order of magnitudes,

eDRAM ~ hundreds of GB/s
CPU-GPU ~ tens of GB/s
XDR RAM ~ tens of GB/s

And on the subject of Anti-aliasing, that's why real-time REYES rendeing has been brought up with the recent Toshiba patent linked and thread using micro-polygons. Which is the reason for the Poll given the flexibilty of the PS3 Cell architecture.

 

[Ailuros]

IMO, when Simon made that comment about Jason's post, it's kinda rude not to give any explanation to the uninitiated...or clarify the post. A simple 'he's clueless' is not constructive, though, I'm sure Simon didn't intend to.. ...So in his absence can you elaborate on what Simon was implying?

No. What makes you think that I'm entitled to speak on his behalf?

Have you read the patent? Also please read my reply to Faf above. I'm not aware of any dealing with 'tiles' and overlapping 'bricks' to render scenes the way that patent describes.

Yes I read it and? You might want to read into SONY's patents refering to their former console designs also or ask around how those exactly where rendering.

Hey, there are patents for those at least and I've already mentioned to prefix 'hybrid' to TBDR at the beginning of the thread and it's also the 'title' of another linked thread! ...If your really interested in bandwidths, these would be my order of magnitudes,

eDRAM ~ hundreds of GB/s
CPU-GPU ~ tens of GB/s
XDR RAM ~ tens of GB/s

And on the subject of Anti-aliasing, that's why realt-time REYES rendeing has been brought up with the recent Toshiba patent linked above using micro-polygons. Which is the reason for the Poll given the flexibilty of the PS3 Cell architecture.

I still don't see any Tile based deferred renderer.

Maybe the following will help you out too: the XBox is rumoured to have =/>55GB/sec memory bandwidth, ~33 from the GPU and ~22 from 10MB eDRAM. Now take a long hard guess it'll most likely employ either ATI's patented Supertiling method or an evolutionary version of it. In short it's an expansion what ATI has been using for generations over generations (tiling in general) and it splits up the scene first in Macro and then into Micro-tiles. Macro-tiles can AFAIK be either divided over multiple chips or multiple units within one single chip. The XBox2 is rumoured to be able to fit 4xMSAA with a method that could look like a filter on scanout similar approach, from within the eDRAM.

Now believe it or not a real TBDR wouldn't need as much bandwidth to start with. TBDR are widely considered as bandwidth saving sollutions, as they are considered so far to be a rather inadequate idea for extremely high polygon count scenes. If anything else dig up the raw specs of the Dreamcast it might help too and no it's no real comparison to PS1 or PS2 either, since it's not only a vastly different architecture but was released in between too.

 

[one]

Yes, it would be nice if nAo could edit his post and correct that as it's been a reoccurring problem in these discussions. Also, as I stated to you (One) in another thread recently, Ohba (this patent guy) was one of the prominent SCE researchers who had input into the Emotion Engine design.

Though I pointed out in the other thread that Toshiba is the main designer of PS2 processors, input from software side (SCE) seems just getting larger and clearer in PS3 rather than vague blueprint. (The recent patents about EE/VU is IIRC taken by SCE, not Toshiba, so SCE's vision was very clear in hardware-wise even in PS2 and it's not strange considering the nature of Sony group - I'm really curious about how Microsoft challenges this STI synergy in Xbox2 by customizing IBM+ATI hardware for its purpose beyond Flipper or the modified CPU/GPU in Xbox.)

Did search by Akio OHBA, and indeed he has interesting patents, though it's hard to determine which one is for PS/Pocket Station/PS2/PSP/PS3 as there are so many... and I noticed that when both Sony and Toshiba engineers appear on inventor's name, applicant's name is just blank, neither SCE nor Toshiba.

 

[passerby]

Trying to guess PS3/Cell stuff is like playing poker in a pitch-black room with blank cards, and a dealer who won't tell you the rules. I've given up sometime ago. I'll just wait for the officical press/info release.

 

[j^aws]

IMO, when Simon made that comment about Jason's post, it's kinda rude not to give any explanation to the uninitiated...or clarify the post. A simple 'he's clueless' is not constructive, though, I'm sure Simon didn't intend to.. ...So in his absence can you elaborate on what Simon was implying?

No. What makes you think that I'm entitled to speak on his behalf?

Do I need a lawyer or what! ...since you poined out that Simon's point wasn't clarified, and your 'keenness', I thought maybe you had an answer? if not speaking on his behalf, can you or anyone else clarify it?

Have you read the patent? Also please read my reply to Faf above. I'm not aware of any dealing with 'tiles' and overlapping 'bricks' to render scenes the way that patent describes.

Yes I read it and? You might want to read into SONY's patents refering to their former console designs also or ask around how those exactly where rendering.

Can you point me to the 'multitude' of current systems that are currently using parallel renderring of overlapping arbritary sized 'bricks' that are z-merged into tiles/quadrants, then said quadrants combined to form the final image? ...I'm not aware of these 'overlapping bricks' being in common use as described in that patent. If anyone does, then please enlighten me!

Hey, there are patents for those at least and I've already mentioned to prefix 'hybrid' to TBDR at the beginning of the thread and it's also the 'title' of another linked thread! ...If your really interested in bandwidths, these would be my order of magnitudes,

eDRAM ~ hundreds of GB/s
CPU-GPU ~ tens of GB/s
XDR RAM ~ tens of GB/s

And on the subject of Anti-aliasing, that's why realt-time REYES rendeing has been brought up with the recent Toshiba patent linked above using micro-polygons. Which is the reason for the Poll given the flexibilty of the PS3 Cell architecture.

I still don't see any Tile based deferred renderer.

Maybe the following will help you out too: the XBox is rumoured to have =/>55GB/sec memory bandwidth, ~33 from the GPU and ~22 from 10MB eDRAM. Now take a long hard guess it'll most likely employ either ATI's patented Supertiling method or an evolutionary version of it. In short it's an expansion what ATI has been using for generations over generations (tiling in general) and it splits up the scene first in Macro and then into Micro-tiles. Macro-tiles can AFAIK be either divided over multiple chips or multiple units within one single chip. The XBox2 is rumoured to be able to fit 4xMSAA with a method that could look like a filter on scanout similar approach, from within the eDRAM.

Now believe it or not a real TBDR wouldn't need as much bandwidth to start with. TBDR are widely considered as bandwidth saving sollutions, as they are considered so far to be a rather inadequate idea for extremely high polygon count scenes. If anything else dig up the raw specs of the Dreamcast it might help too and no it's no real comparison to PS1 or PS2 either, since it's not only a vastly different architecture but was released in between too.

I'm not calling it a TBDR, it's like the tenth time in this thread that I've said 'hybrid'! ...Semantics aside, my interest lies in that patent (which also describes velocity based LOD and distributed rendering etc.) being realized for PS3 and the point of the Poll is if the architecture is flexible enough to provide in addition to that, Real-time REYES rendering/ Real-time GI rendering etc..

And yes, I was aware of ATI using tiling, but I'm not aware of it using the SCE patents object based 'bricks'/groups' that are overlapping, being rendered/shaded independently from each other with different techniques etc..etc.. A link to their 'supertiling patent' would be useful for comparision? ...Personally I'm hoping for REYES to deal with anti-aliasing using micro-polygons for PS3. 8)

I'm aware of TBDR being efficent with bandwidth and fillrate. But I see it both ways. If they are high values, it doesn't necessarily suggest the contrary, it could mean you get even more performance from the system!

 

[jvd]

Would you like me to edit your post topic then ?

Anyway i remember all the hype from the ps2 and none of it came true. So call me skeptical . But i highly doubt any of what your saying is going to happen .

3.6 ghz cell chip ? Heh yea right what year are they going to make that . The longer they wait to launch the worse a position they are in.

If they launch in 2006 xbox 2 will have a year head start over them so even if the ps3 has a slight edge in what it can put out on the screen it will hardly matter (just like the xbox) If they launch in 2007 , well thats 2 years and 2 x mases that the xbox 2 had to its self and ms could in theory launch a xbox 3 only 2 short years after which will put them in a crappy situtation again . Anything after 2007 and they are better off just skipping this gen

 

[Vince]

3.6 ghz cell chip ? Heh yea right what year are they going to make that

Just so you know, many of the lead members on Cell development were previously working on BOA and guTS (both high-speed architecture projects) and having finished Cell work are now tasked with ultra-high speed future-generation PowerPC products. Don't speak to soon.

 

[jvd]

3.6 ghz cell chip ? Heh yea right what year are they going to make that

Just so you know, many of the lead members on Cell development were previously working on BOA and guTS (both high-speed architecture projects) and having finished Cell work are now tasked with ultra-high speed future-generation PowerPC products. Don't speak to soon.

Yea thankfully i'm not speaking to soon.

I like how when you attach sony to something , there are people in this forum who will believe it .

watch after the ps3 comes out we are gonig to start hearing ps4 in 2011 200 tflops power . Heh.

As i said look at ps2 hype and look what it turned out to be. Same will happen here.

 

[Vince]

As i said look at ps2 hype and look what it turned out to be. Same will happen here.

And what did it turn out to be? A machine that dispite it's nearly 2 year disadvantage against the competition can keep within distance of it on a graphical level? That the architecture, when related to it's timeframe and technology available, is superior than the competition? The hype was right, for a 1999 machine it kicks ass.

I really don't see why it's your mission to set the record strait when there is no discrepency to set right.

 

[jvd]

As i said look at ps2 hype and look what it turned out to be. Same will happen here.

And what did it turn out to be? A machine that dispite it's nearly 2 year disadvantage against the competition can keep within distance of it on a graphical level? That the architecture, when related to it's timeframe and technology available, is superior than the competition? The hype was right, for a 1999 machine it kicks ass.

I really don't see why it's your mission to set the record strait when there is no discrepency to set right.

what it turned out to be was a disapointment ? where is the 66 million polygons i should be seeing ? Where is the power of the system we should have passed up dc for ? Even now dispite its year advantage of price premium its games are not a lvl beyond dc games. Best case example is , playstation graphics compared to saturn graphics . Yes you can argue that it holds its own against the xbox (though i can hardly say so just byl ooking at gta 3) but when you compare it to a gc which cost less to manufacture it doesn't cast the ps2 in such great light


YOu like to dream big and thats great. But list it as dreaming.

Whats to stop me from dreaming ?

Oh the ms system is going to have a 64x2 pipe line set up that can operate like a 128 pipeline set up and run at 1ghz ? What impossible u say ? Well thats what they said about getting a part above 300mhz on 150nm process and those guys did it !

Or i can say well Heh a cell chip at 3.6 ghz ? the chip in the ms system will be 18ghz with 6 cores . Don't believe it well guess what they are partialy responsible for your cell chip thats much more complicated and is going to hit 3.6 ghz ?

See what I"m getting at?

You want me to believe that a chip infinitly more advance than a p4 is going to scale higher than a p4 is currently scaling ?

The p4 on 90nm is having trouble past speeds of 3.6 ghz and you want me to bleieve that a 65nm part that is much more complicated and much bigger is going to hit 3.6 ghz ?Not to mention it will have to make due with less efficent cooling all around , heatsink wise , air flow into and out of the case and fan power on the actual heatsink. You haven't shown any proof of it either, you just say well trust me . Who are you to trust ?


If you want me to believe that show me a chip close in complexity to the cell chpi that is hitting close to 3 ghz .

THe last solid info i had places the cell at 2.8ghz for the cell chip they want in the ps3 and guess what that is if they are able to launch with 45nm tech.

Last solid info i had was 256mbs of yellowstone ram and 64mbs in total of on die sram. 32 in each chip .

This data i got was 2 weeks ago and its from a solid source .

The only thing in that diargram that is corret is the clock speed of the graphics portion which will in fact be 800mhz (Desired speed)

 

[pahcman]

And what did it turn out to be? A machine that dispite it's nearly 2 year disadvantage against the competition can keep within distance of it on a graphical level? That the architecture, when related to it's timeframe and technology available, is superior than the competition? The hype was right, for a 1999 machine it kicks ass.

"keeping within distance" is a nice way of putting it. The factual truth is, its "holding back competition". ~70m units sold is a powerful trump card for being the base of development, as clearly platform specific games, genres to genres, showed clearly the difference.

I dont know whats your definition of "architecture superiority" is, the way i see it, broken mipmaps, lower resolution, poor/no detailed textures, no pixel blending, hardly can i call something like that kick ass for 2000.

Maybe you view graphics differently, but for me, those are as important as any other factors. When judging graphics technology, I like the whole package to be presentable for its time, not just parts by parts.

 

[DeanoC]

First I have no idea how that patent discribes a tile based defered renderer. If thats a TBDR than so is ATI...

Also any system can do REYES (certainly any next-gen console will probably have a GPU capable of REYES). The question is wether the visual results are worth the gains, thats almost impossible to answer without things like memory latency and pipeline structures...

REYES removes pixel shading by using very small triangles, this can be already done today but isn't because losing a coherance variable is expensive. The lose of fillrate coherance (quads), on a current PC this would cost alot of fillrate. Unless PS3 rasterisor has completely seperate single pixel pipes than REYES could waste fillrate like nobodies business.

Ignoring that issue, you can see unified shader architectures as a partial move to REYES (you can choose to switch 'off' the pixel shaders) and then its just a question of the tesselation and triangle rate.
For a 640x480 with 1 quad per pixel, you need 18.5 million quads per depth layer per second. For 720p you need 55.2 Million quads . If you expecting HDTV res and REYES you going to need sustained quad rate of at least a billion per second (not impossible but not exactly easy).


GI is both a math and database problem. Cell seems to be the antithesis of fast database searches (which generally require lots of RAM, not small pools). So I don't see any major use of GI coming to a PS3 (maybe first bounce approximation like NVIDIA demo on GFFX).

In conclusion I'd say that a REYES style renderer is possible, if they design the hardware that way (only the future will tell) but GI isn't happening in any major way.

 

[nAo]

Obviously that patent does not describe any TBDR architecture, it's pretty clear it's an immediate mode renderer.
As DeanoC said going to an unified shader architecture is already a 'little' step forward to REYES but, imho, it's not enough in any way to make realtime REYES a big deal.
So even if REYES will be possible on Xenon (well..it even possible on the PS2 ) at the moment I can't see any reason why one would want to take the REYES route on a standard architecture.
But REYES is more than rasterizing a micropolygons grid and an architecture designed around REYES would be fresh air in this industry.

ciao,
Marco

 

[Fafalada]

Maybe I missed something from the patent but I see both Tiles and Bricks.

I don't - there's no mention of tiles aside for that one lone comment about non-overlapping bricks looking similar to them. While using TBDR in combination with what is described in patent is perfectly possible, the patent itself never suggests it.

The lose of fillrate coherance (quads), on a current PC this would cost alot of fillrate. Unless PS3 rasterisor has completely seperate single pixel pipes than REYES could waste fillrate like nobodies business.

Well I've mentioned this before - the hardware should be designed around Reyes pipeline for it to be efficient. There's no point in wasting transistor on pixel pipelines that can texture if the objective is to not use them...
I don't think the rasterizer part would be the biggest problem really, it would only need to do very simply processing. The tesselation/primitive processing efficiency would be more troublesome I think, especially if we really used programmable units for it (and it would really need to be programmable for this to work well).
Actually I'm more worried about getting some halfway solution that doesn't do anything particularly well.

As for GI, to stay on topic, I think we'll have to make do with PRT like stuff for the time being.

 

[MfA]

Obviously that patent does not describe any TBDR architecture, it's pretty clear it's an immediate mode renderer.

Oh hell no, regardless of overlap it does world space selection of groups of objects which are localized in space (both 3D and 2D obviously) and renders those groups seperately. That is not immediate mode rendering ... and as I said, I think a system based on that patent would be operating in this mode 99% of the time :

"Thus, when the space is divided so that there are no overlapping bricks in the x,y directions, it is possible to do rendering processing similar to tiling and generate a large size final image using small high-speed VRAM area."

 

[nAo]

Oh hell no, regardless of overlap it does world space selection of groups of objects which are localized in space (both 3D and 2D obviously) and renders those groups seperately. That is not immediate mode rendering ...

You're correct if we assume that all the selection work is done by the hardware, but I don't think that will be the case in the final hw. (if it will ever exist)
Each selected block (or brick) is rendered in immediate mode, IIRC.

ciao,
Marco

 

[Fafalada]

Oh hell no, regardless of overlap it does world space selection of groups of objects and renders those groups in parallel. That is not immediate mode rendering ...

Er - they mentiond doing the grouping, but never explain how Going by my experience with Sony that probably translates as "do it yourself!".
That aside, PS2 is used as deferred renderer 99% of the time too - but it doesn't get called that because of rasterizer being "immediate" mode.

and as I said, I think a system based on that patent would be operating in this mode 99% of the time :

Some would argue that GSCube begs to differ
Anyway, my view of idea of rendering in layers is to have self contained groups of data - near perfect locality, and with it paralelization.
Tiling non-overlapping groups in screen space will have massive data dependancies/duplication across groups, and perhaps more importantly, plenty of workload duplication.
Both have tradeoffs obviously, but I prefer the former myself.

 

[Tuttle]

The problem with extrapolating from the various patent documents is that, as far as I know, Sony plans a wide range of computing devices built upon the technology. There is no way of knowing what applies to the ~$300ish PS3s, ~$3000 to $10000ish workstations, or other as of yet unannounced devices.

 

[MfA]

Anyway, my view of idea of rendering in layers is to have self contained groups of data - near perfect locality

Localization in some arbitrary layer in 3D space hits the framebuffer, textures and geometry as randomly as standard rendering (actually more so in case of grouping by state). The only thing it helps with is front to back rendering.

Parallelization by rendering in layers and doing back end compositing is stupid. If you dont want to benefit from tiling you might as well use sort middle and dont have the need to have N framebuffers before compositing. This is AFAIK what ATI uses at the moment for instance.

 

[Fafalada]

Parallelization by rendering in layers and doing back end compositing is stupid anyway (BTW it mentions tiling as an option in my quote, the same can not be said for rendering in layers).

I think we have a terminology confusion - what I refer to as layer is equivalent to their brick - it has depth information. And obviously, they don't have to be fullscreen sized, although it makes some things simpler if they are.

Won't argue which term is better suited, but I've seen layers used in that manner for much longer then bricks