rendezvous
Regular
Urian said:
No, raytracing is all about vectors.
nVidia shader patent (REYES/Raytracing/GI) destined for PS3?
- Thread starter j^aws
- Start date
No, raytracing is all about vectors.
Fox5 said:
That would need a single common architecture. Or else each developer would have to program their application for each and every different configuration of PC there is out there, and there are LOTS of different configs...
Also, how would you manage all your programs?
There has to be an OS, even a very simple one, but there needs to be one.
I'm not exactly sure why Windows is now as bloated as it is, how it happened and why it takes something like 10-20GB although coming from a simple 650MB CD, but no OS at all aint gonna work.
Now even mobile phones have OSs, because it's easier for the end user to manage a device if there is a program that takes care of all the stuff "in the background", and OS.
london-boy said:Fox5 said:
That would need a single common architecture. Or else each developer would have to program their application for each and every different configuration of PC there is out there, and there are LOTS of different configs...
Also, how would you manage all your programs?
There has to be an OS, even a very simple one, but there needs to be one.
I'm not exactly sure why Windows is now as bloated as it is, how it happened and why it takes something like 10-20GB although coming from a simple 650MB CD, but no OS at all aint gonna work.
Now even mobile phones have OSs, because it's easier for the end user to manage a device if there is a program that takes care of all the stuff "in the background", and OS.
Single common architecture = playstation 3.
And maybe it would make you put in a disk when you want to use the application, or they could have a simple menu from where you could select the application you want to use.
Fox5 said:
But that's not a PC then, that's still a console. I might want to use more than one application at once. And also it would be nice not to have to insert a disc everytime i need the application to start.
An OS will always be needed. Shame that we only have Windows now, and it doesn't give people much of a choice or even a nice idea of OS's in general.
a688 said:
:| No whay!!!??? Their r moer OSs ohter then Windoz!!11!!????? And u meen taht My Docuemnts aint included in the figurez???
Obviously my post wasn't clear enough. I apologise.
Given that half of that "breadth of functions" is crap that makes the system worse then it could have been, I don't see why not.pahcman said:
Now if you said Windows 2000, I would be more inclined to agree with you.
Of course arguing about OS functionality is all fluff - what defines a successfull OS is the software support (not unlike consoles
) and that's where a new born OS will obviously always have an uphill battle.CELL + nVidia = shaded RTRT potential
In re-reading this paper on ray tracing recently:
They mention that they can do real basic ray tracing with an FPGA they programmed running at 4 Gflops. However, they also mention:
Wouldn't the CELL + nVidia ROP provide the essential combination for this to happen?
Could this be why they seemingly won't use nVidia for the whole graphics pipeline so that they can leave this road open?
In re-reading this paper on ray tracing recently:
They mention that they can do real basic ray tracing with an FPGA they programmed running at 4 Gflops. However, they also mention:
Wouldn't the CELL + nVidia ROP provide the essential combination for this to happen?
Could this be why they seemingly won't use nVidia for the whole graphics pipeline so that they can leave this road open?
Re: CELL + nVidia = shaded RTRT potential
Maybe an advanved RT implementation on CELL could exploit rays coherence thus grouping queries on some spatial subdivision structure to a void a lot of scattered memory read operations, otherwise A|SPUs built-in dma prefetch mechanism would be uneffective!
Once DeanoC wrote that a future PS4 and XBOX3 comparison would be based on external memory latencies and not raw flop/s figures..I couldn't agree more
One key performance figure I would like to know about A|SPUs is the time needed to switch thread and it would be nice to have some kind of support for fast registers saving and restoring, or fast registers banks switching.
A|SPUs flexibility is a given..one can do everything, the real problem is how to do everything keeping A|SPUs don't run idle most of the time..
Your question is not unexpcted anyway, at some point GPU manufacters would have to introduce some very fast and flexible calculations core...maybe something along the lines of CELL architecture could be one of the answers and NVIDIA would like to use it.
Too bad I think that's not the case..I'm still skeptic about CELL stuff being into PS3 GPU
ciao,
Marco
APUs local memory is so small that one can't just pretend to run a plain RT implementation on them without issuing a lot of external memory references via DMA requests. That's the kind of things one doesn't want to happen on a stream processor.PZ said:
Maybe an advanved RT implementation on CELL could exploit rays coherence thus grouping queries on some spatial subdivision structure to a void a lot of scattered memory read operations, otherwise A|SPUs built-in dma prefetch mechanism would be uneffective!
Once DeanoC wrote that a future PS4 and XBOX3 comparison would be based on external memory latencies and not raw flop/s figures..I couldn't agree more
One key performance figure I would like to know about A|SPUs is the time needed to switch thread and it would be nice to have some kind of support for fast registers saving and restoring, or fast registers banks switching.
A|SPUs flexibility is a given..one can do everything, the real problem is how to do everything keeping A|SPUs don't run idle most of the time..
Why are you saying they seemingly won't use nvidia IP for the whole pipeline? I don't think we have any data about or am I missing something here?
Your question is not unexpcted anyway, at some point GPU manufacters would have to introduce some very fast and flexible calculations core...maybe something along the lines of CELL architecture could be one of the answers and NVIDIA would like to use it.
Too bad I think that's not the case..I'm still skeptic about CELL stuff being into PS3 GPU
ciao,
Marco
Since CELL is a stream processor, here's an interesting ppt on how to do raytracing on a stream processor by mapping to a GPUs programmable fragment processor...
Ray Tracing on Programmable Graphics Hardware ppt.
So could the NV GPU just be one big fragment processor complementing the CELLs programmable vertex processor
Ray Tracing on Programmable Graphics Hardware ppt.
So could the NV GPU just be one big fragment processor complementing the CELLs programmable vertex processor
Jaws said:Since CELL is a stream processor, here's an interesting ppt on how to do raytracing on a stream processor by mapping to a GPUs programmable fragment processor...
Ray Tracing on Programmable Graphics Hardware ppt.
So could the NV GPU just be one big fragment processor complementing the CELLs programmable vertex processor
Are you sure about the link?
I only get "The page cannot be displayed".
pc999 said:Jaws said:Since CELL is a stream processor, here's an interesting ppt on how to do raytracing on a stream processor by mapping to a GPUs programmable fragment processor...
Ray Tracing on Programmable Graphics Hardware ppt.
So could the NV GPU just be one big fragment processor complementing the CELLs programmable vertex processor
Are you sure about the link?
I only get "The page cannot be displayed".
http://online.cs.nps.navy.mil/Dista...g/2002/Papers/13_GraphicsHardware/purcell.ppt
It works for me...it's a direct link...PM me if it still doesn't work...
Re: CELL + nVidia = shaded RTRT potential
Yes you are right in that the APU memory is limited. But I am not ready to give up yet. Here are some reasons supporting the potential for a ray tracing form of rendering with a CELL + nVidia architecture:
1) As Jaws mentioned, GPU ray tracers have been around now as a proof of concept of ray tracing on a streaming processor. They also highlight just how wonky it is to use GPUs as ray tracing engines. William Dally talking about the Imagine Stream Processor from Stanford also explicitely mentioned ray tracing in a streaming architecture. I agree it is not that great a match for RT but it is a potential avenue.
2) There are opportunities for coherence in rays especially for rays with common origins (shooting multiple shadow rays at a point to multiple lights) or rays with common directions.
3) Hofstee in his video mentions global illumination (as I recall)
4) The paper I linked to above also includes some discussion about memory bandwidth and access patterns. They indicate that ray tracing does require pretty random access patterns but that this keep different banks of memory utilized fully instead of everyone clobbering one chunk.
Their own ray tracing pipeline requires 75 KB of memory per pipeline which could fit in an APU memory plus an overall cache to prevent triangles from being intersected multiple times. They have 64 threads going on at a time. They estimate bandwidth (worst case) of 2GB/s bandwidth for 1024x768x60Hz which is easy for Cell.
They mention that their 90Mhz FPGA is equivalent to an 8-12Ghz CPU and is about 3-5x GPU implementation because of its special purpose design. The Cell is not this special purpose but does have part of the streaming advantage along with the clockspeed advantage. As you keenly mentioned, the critical element is the special bits they added for keeping track of traced objects and rebuilding ray tracing object hierarchies as objects move. These kinds of things that are special purpose to their chip and which could not be efficiently ported over are the unknowns. If these unkowns are critical then yes the dream is over (for now).
Ray tracing is not sooo bad because it is completely demand driven. You are only processing pixels/triangles that contribute to the final scene and in that sense you are only asking of the memory what is relevant.
Yes I think if you look at the supercomputer market you can see that Cray has made advances not just because of pure speed but because their data transfer systems in the machine are low latency (the lowest around).
I think also that Longhorn has some specific requirements about rendering latency and handling huge numbers of threads (1000s). I believe they want every icon to be in 3D and in a separate thread. Stuff like that must cause nVidia and ATI a lot of headaches.
I am just taking a hint from the patent which shows the Cell top side and Pixel Engine bottom side as well as an image from a patent (not sure which one) which shows the APUs set up as a vertex processing pipeline. Not an airtight case by any means but it's all we have so far.
nAo said:
Yes you are right in that the APU memory is limited. But I am not ready to give up yet. Here are some reasons supporting the potential for a ray tracing form of rendering with a CELL + nVidia architecture:
1) As Jaws mentioned, GPU ray tracers have been around now as a proof of concept of ray tracing on a streaming processor. They also highlight just how wonky it is to use GPUs as ray tracing engines. William Dally talking about the Imagine Stream Processor from Stanford also explicitely mentioned ray tracing in a streaming architecture. I agree it is not that great a match for RT but it is a potential avenue.
2) There are opportunities for coherence in rays especially for rays with common origins (shooting multiple shadow rays at a point to multiple lights) or rays with common directions.
3) Hofstee in his video mentions global illumination (as I recall)
4) The paper I linked to above also includes some discussion about memory bandwidth and access patterns. They indicate that ray tracing does require pretty random access patterns but that this keep different banks of memory utilized fully instead of everyone clobbering one chunk.
Their own ray tracing pipeline requires 75 KB of memory per pipeline which could fit in an APU memory plus an overall cache to prevent triangles from being intersected multiple times. They have 64 threads going on at a time. They estimate bandwidth (worst case) of 2GB/s bandwidth for 1024x768x60Hz which is easy for Cell.
They mention that their 90Mhz FPGA is equivalent to an 8-12Ghz CPU and is about 3-5x GPU implementation because of its special purpose design. The Cell is not this special purpose but does have part of the streaming advantage along with the clockspeed advantage. As you keenly mentioned, the critical element is the special bits they added for keeping track of traced objects and rebuilding ray tracing object hierarchies as objects move. These kinds of things that are special purpose to their chip and which could not be efficiently ported over are the unknowns. If these unkowns are critical then yes the dream is over (for now).
Ray tracing is not sooo bad because it is completely demand driven. You are only processing pixels/triangles that contribute to the final scene and in that sense you are only asking of the memory what is relevant.
Once DeanoC wrote that a future PS4 and XBOX3 comparison would be based on external memory latencies and not raw flop/s figures..I couldn't agree more
One key performance figure I would like to know about A|SPUs is the time needed to switch thread and it would be nice to have some kind of support for fast registers saving and restoring, or fast registers banks switching.
A|SPUs flexibility is a given..one can do everything, the real problem is how to do everything keeping A|SPUs don't run idle most of the time..
Yes I think if you look at the supercomputer market you can see that Cray has made advances not just because of pure speed but because their data transfer systems in the machine are low latency (the lowest around).
I think also that Longhorn has some specific requirements about rendering latency and handling huge numbers of threads (1000s). I believe they want every icon to be in 3D and in a separate thread. Stuff like that must cause nVidia and ATI a lot of headaches.
Could this be why they seemingly won't use nVidia for the whole graphics pipeline so that they can leave this road open?
Why are you saying they seemingly won't use nvidia IP for the whole pipeline? I don't think we have any data about or am I missing something here?
Your question is not unexpcted anyway, at some point GPU manufacters would have to introduce some very fast and flexible calculations core...maybe something along the lines of CELL architecture could be one of the answers and NVIDIA would like to use it.
Too bad I think that's not the case..I'm still skeptic about CELL stuff being into PS3 GPU
ciao,
Marco
I am just taking a hint from the patent which shows the Cell top side and Pixel Engine bottom side as well as an image from a patent (not sure which one) which shows the APUs set up as a vertex processing pipeline. Not an airtight case by any means but it's all we have so far.
Jaws said:
I got it here if anyone wants
http://cggmwww.csie.nctu.edu.tw/sem...tracing.on.programmable.graphics.hardware.ppt
Interesting R8500 is better than GF3, I am right?
Raytracing on GPUs seem like a dead end to me.
The benefit of raytracing is in the way it allows a more correct handling of refraction and specular reflection. This makes it ill suited for a stream based approach because it introduce a data dependent step that has, potentially, highly irregular access patterns (cache busting).
I can see GPUs as accelerating first intersection rays (which is also what the above paper deals with), but these account for just a fraction of the work a heavy duty raytracer has to do.
Cheers
Gubbi
The benefit of raytracing is in the way it allows a more correct handling of refraction and specular reflection. This makes it ill suited for a stream based approach because it introduce a data dependent step that has, potentially, highly irregular access patterns (cache busting).
I can see GPUs as accelerating first intersection rays (which is also what the above paper deals with), but these account for just a fraction of the work a heavy duty raytracer has to do.
Cheers
Gubbi
Similar threads
