Return to rasterizer for PS4?

[nAo]

Actually it does, sounds like revisionist history to me. I don't think the patent mentioned the Broadband Engine, Visualizer and CELL for nothing.

Too bad that patent was NOT about visualizer.
Sony, IBM and Toshiba have tons of patents about GPUs and none of them are based on CELL, none, nada, zero, nil, niente di niente.
The fact is a CELL based GPU has never existed, not even on a drawing board for more than 10 seconds.

 

[Capeta]

Too bad that patent was NOT about visualizer.
Sony, IBM and Toshiba have tons of patents about GPUs and none of them are based on CELL, none, nada, zero, nil, niente di niente.
The fact is a CELL based GPU has never existed, not even on a drawing board for more than 10 seconds.

No? Then how come that diagram is titled:

Figure 6 from US Patent 2002/0138637 - Suzuoki, Masakazu and Yamazaki, Takeshi
Broadband Engine and Visualizer, both leveraging "CELL"

and...

Figure 8 from US Patent 2002/0138637 - Suzuoki, Masakazu and Yamazaki, Takeshi
Single Chip Broadband Engine and Visualizer

And why does the Visualizer look eerily similiar to the CELL architecture?

PU=PE
APU=SPE

[0013] The basic processing module is a processor element (PE). A PE preferably comprises a processing unit (PU), a direct memory access controller (DMAC) and a plurality of attached processing units (APUs). In a preferred embodiment, a PE comprises eight APUs. The PU and the APUs interact with a shared dynamic random access memory (DRAM) preferably having a cross-bar architecture. The PU schedules and orchestrates the processing of data and applications by the APUs. The APUs perform this processing in a parallel and independent manner. The DMAC controls accesses by the PU and the APUs to the data and applications stored in the shared DRAM.

Coincedence or revisionist history?

 

[Carl B]

nAo was at the leading edge of following the STI-related graphics patents back in the day, so I assure you there's no need to "clear things up for him" as far as the patent goes.

What he's saying (and nAo will correct me if I'm wrong) is that the patent is about Cell, not about the Visualizer - the Visualizer was mentioned as a placeholder of sorts in that patent simply because configuring a Cell like that would be an option, and it's something they thought of when they drafted it. Hell when you get down to it the patent isn't even about Cell per se, but an overarching concept of the 'Broadband Engine' architecture as a whole, of which the modern Cell is a subset.

What people are trying to point in any number of these PS3 graphics retrospective threads lately is that when it came down to actual R&D engineering efforts at Sony, there was a GPU design being co-developed with Toshiba, but that design was *not* Cell-based and it was *not* patterned on the Visualizer concept.

 

[Capeta]

What he's saying (and nAo will correct me if I'm wrong) is that the patent is about Cell, not about the Visualizer - the Visualizer was mentioned as a placeholder of sorts in that patent simply because configuring a Cell like that would be an option, and it's something they thought of when they drafted it. Hell when you get down to it the patent isn't even about Cell per se, but an overarching concept of the 'Broadband Engine' architecture as a whole, of which the modern Cell is a subset.

What people are trying to point in any number of these PS3 graphics retrospective threads lately is that when it came down to actual R&D engineering efforts at Sony, there was a GPU design being co-developed with Toshiba, but that design was *not* Cell-based and it was *not* patterned on the Visualizer concept.

Sure if you want to put it that way then yeah the "Visualizer" term may have been used as a "placeholder" but there is zero doubt that those diagrams was referring to a "preferred embodiment" of a CELL architecture. I wasn't trying to say that diagram of the "Visualizer" was what the actual GPU that Toshiba designed turned out to be or if it was even based on a CELL architecture.

 

[Carl B]

Sure if you want to put it that way then yeah the "Visualizer" term may have been used as a "placeholder" but there is zero doubt that those diagrams was referring to a "preferred embodiment" of a CELL architecture.

Sure, but just so we're clear "preferred embodiments" don't relate directly to what the patent authors actually intend to do per se. Preffered embodiment descriptions are used to basically wrap up every single idea that you can into a patent, such that if anyone infringes on it in the future, even if you actually didn't put those ideas into practice yourself, there you are - protected and with the patent. I just wanted to clarify that aspect because you're highlighting 'revisionist history' in terms of the Broadband Engine patent, but quite on the contrary, many of us are quite familiar with said patent and understand what you mean quite well when you mention "Visualizer."

I think the tension here comes from approaching the situation from opposite directions: your wanting to highlight a change from what was written in the original patent, and in this case nAo wanting to emphasize that what was in the patent was never relevant to actual development. In my mind, these are not mutually exclusive.

I wasn't trying to say that diagram of the "Visualizer" was what the actual GPU that Toshiba designed turned out to be or if it was even based on a CELL architecture.

Sure, understood. Just so long as you understand though that what Toshiba was working on was never based on Visualizer to begin with.

 

[Acert93]

64 SPEs in total should be enough to do some decent ray-tracing.

Will it?

Will a 64 SPE configuration outpace the results of a rasterizer? (Ignoring the fact SPEs will be used for tasks other than graphics).

While 5 years is a long time, I think there is a lot of movement the other way, namely GPUs taking on more tasks. As much as we are looking at CPUs taking on graphic tasks, the reverse is happening (possibly even faster) with tasks like physics already being offloaded to the GPUs.

I wrote a while back that we may see MS and Sony go different routes. I wouldn't be surprised to see a PS4 that is Cell-centric and a Xbox3 that is GPU centric with silicon budgets representing the divergent paths to tackling simular issues.

 

[Crossbar]

64 SPEs in total should be enough to do some decent ray-tracing.

Will it?.

I think so.

Will a 64 SPE configuration outpace the results of a rasterizer? (Ignoring the fact SPEs will be used for tasks other than graphics).

I think that is a somewhat other question.

While 5 years is a long time, I think there is a lot of movement the other way, namely GPUs taking on more tasks. As much as we are looking at CPUs taking on graphic tasks, the reverse is happening (possibly even faster) with tasks like physics already being offloaded to the GPUs.

I wrote a while back that we may see MS and Sony go different routes. I wouldn't be surprised to see a PS4 that is Cell-centric and a Xbox3 that is GPU centric with silicon budgets representing the divergent paths to tackling simular issues.

That may very well be the case where MS will encapuslate the GPU with APIs providing powerful graphics and physics functionality and Sony will provide a more down to the metal approach via Cell (and possibly another nvidia GPU).

 

[Acert93]

I think so.

Ok, based on what though? Cell has shown to be a good raytracer, but it still seems to be a bit away from providing the results we are seeing in current GPUs, let alone GPUs of the 2010/2012 timeframe.

I think that is a somewhat other question.

Not IMO.

RT needs to be at least roughly comparable to rasterizing to merit such a transition. If a RT solution has 1/2 the result for the same silicon budget compared to a rasterizer solution then the RT solution then it probably isn't decent enough.

Being "decent enough" is all relative to the competition. If you are spending the same amount of money on hardware but getting back a substantially less return in results then in relative terms it isn't decent.

In theory some day RT will catch up and pass rasterizing. Until then rasterizing isn't standing still; a RT solution not only needs to exceed the results of today's GPUs but those of tomorrow (plus all the software/legacy/development issues).

I am hopefully for RT down the road as it may offer some simplicity and refocus toward design. But it really needs to give comparable results on screen before it becomes really interesting for the gaming space.

 

[Shifty Geezer]

I think so.

I think not It's the memory accessing that holds RT back. Cell can raytrace in realtime a landscape created on the fly, but put a 3D scene in there with lots of 3D models occluding each other and textures to look up, and suddenly the Cell is spending all it's time digging through memory to test ray intersections and look-up textures. A need for lots of flotaing-point maths power is only part of the limits in raytracing. There's other hurdles that need to be overcome too.

 

[nAo]

No? Then how come that diagram is titled:

Figure 6 from US Patent 2002/0138637 - Suzuoki, Masakazu and Yamazaki, Takeshi
Broadband Engine and Visualizer, both leveraging "CELL"

Have you read what I wrote? STI issued dozens and dozens of patents about GPUs, while they have zero patents about a GPU CELL based and you completely ignore this in favour of 3 lines of text quoted in very very early patent (look at the date) that is not even about a GPU!
Also use the search function and try to guess who was the guy who has found most of this stuff *first*, you're lecturing the wrong person
[thanks Carl ]

 

[Lazy8s]

Considering Sony revealed that a CELL based GPU was one of the ideas they considered, people speculating so much about the potential it could've had is only natural.

 

[Crossbar]

Ok, based on what though? Cell has shown to be a good raytracer, but it still seems to be a bit away from providing the results we are seeing in current GPUs, let alone GPUs of the 2010/2012 timeframe..

I was thinking of the actual ray tracing part not the rendering (the term ray-tracing is a bit ambiguous IMO), I think that 64 Spes providing possibly 2 TFlops of performance will be able to back-trace quite a few rays.

In my original post I made a few other assumptions about changes in the Cell which more concerns the actual rendering of the image, the "etc."-part may include what you are missing:

Two 32 SPE-Cells, where the SPE instruction set has been expanded with some graphics specific instructions and some ROP-units, etc have been added.

I agree with what you are saying that there may be other techniques that provide better images/transistor than ray-tracing. I don´t hold ray-tracing as a holy grale or anything, though it can be used to generate some nice images. :smile:

 

[Crossbar]

I think not It's the memory accessing that holds RT back. Cell can raytrace in realtime a landscape created on the fly, but put a 3D scene in there with lots of 3D models occluding each other and textures to look up, and suddenly the Cell is spending all it's time digging through memory to test ray intersections and look-up textures. A need for lots of flotaing-point maths power is only part of the limits in raytracing. There's other hurdles that need to be overcome too.

I think I understand what you are saying, but I also think there will be algorithms that will help by simplifying objects as well as truncating some of the actuall ray-tracing. Whether you still will call that ray-tracing when you don't do it in a complete fashion true to the actual defintion of ray-tracing, I don´t know.

And I do believe that the next generation will have some insane memory bandwidth compared to the current generation, Samsung already has some impressive GDDR4 at 4 GHzup and running.

 

[Gubbi]

It's not a bandwidth problem, it's a latency/outstanding request problem.

In raytracing you're chasing through space decomposition structures all the time. Coherence of primary rays means that these structures caches well and hence is reasonable fast... But still many times slower than rasterization for the same net visual effect.

Secondary rays is what sets raytracing apart. However coherence of secondary rays is usually (for all interesting cases ) a lot worse, so you end up going to main memory a lot more when processing each node in your space decomposition structure. Going to main memory increases latency by two decimal orders of magnitude. To not idle while waiting for main memory you need 100 times more ray contexts. These contexts take up space (on die).

On top of that your memory system needs to be able to handle a mind-boggling number of independent (poor coherence) memory requests.

Cheers

 

[Megadrive1988]

Sure, but just so we're clear "preferred embodiments" don't relate directly to what the patent authors actually intend to do per se. Preffered embodiment descriptions are used to basically wrap up every single idea that you can into a patent, such that if anyone infringes on it in the future, even if you actually didn't put those ideas into practice yourself, there you are - protected and with the patent. I just wanted to clarify that aspect because you're highlighting 'revisionist history' in terms of the Broadband Engine patent, but quite on the contrary, many of us are quite familiar with said patent and understand what you mean quite well when you mention "Visualizer."

I think the tension here comes from approaching the situation from opposite directions: your wanting to highlight a change from what was written in the original patent, and in this case nAo wanting to emphasize that what was in the patent was never relevant to actual development. In my mind, these are not mutually exclusive.



Sure, understood. Just so long as you understand though that what Toshiba was working on was never based on Visualizer to begin with.

there still seems to be much confusion about the in-development GPUs and hypothetical GPUs Sony would use for PS3, pre-RSX.

the Graphics Synthesizer 3
the supposed CELL-based Visualizer which now seems not to have existed
the Toshiba GPU

I admit i'm still quite confused over it all. it's water under the bridge now though, since PS3 is on the market.

we now have PSP2 and PS4 GPUs to speculate on ^__^

 

[LunchBox]

if it's for a GPU that will do realtime raytracing, I don't know if 64 SPEs would suffice...

 

[Megadrive1988]

Actual realtime raytracing of decent quality with complex scenes & models is going to require a completely new architecture that no current or upcoming known CPU or GPU has.

even in CG, raytracing is used only sparingly since it's so time consuming to render offline.