NVIDIA Patent: G70/G80/RSX?

[j^aws]

Sounds to me like G70 pipes where the texture units are tied in with the shader units, so you either texture or shade but not both, the idea being that if you're not texturing, a specialist texture-unit is silicon sitting idle. Multithreading is reminiscent of Xenos' threading to overcome stalls. I don't know how GPU's handle that already. I guess the idea is if a shader needs a texture read and the texture isn't local, rather than stall the shader pipeline waiting for the fetch, the pipe (processor. quad?) switches switches to another shader thread. That's going solely by the abstract Jaws posted and no investigation of the actual patent, mind!

Well, seems like a unified shader pipeline judging from fig 2. (as version posted in the other thread)

You can see the execution piplelines consuming vertex/fragment buffers. It looks like a unified shader 'quad' with a 'coupled' texture unit... certainly different from Xenos' non-quad, 16 ALU unified shader units with fully decoupled texture units...

 

[RavenFox]

Well maybe we should've used you guys instead!

But I live in DC so... I mean it only makes sense for us to have used a DC firm.

We have a DC office also. I think our firm handled some x-box stuff in the past. Probably the game case designs hmm or was it the cube lol. Sometimes where under NDA's to lol.
I work in the IT dept by the way.

 

[Jawed]

Fig.2 in the patent is very much a unified shading engine with vertex and pixel shading being performed by common execution units.

Jawed

 

[Shifty Geezer]

That diagram definitely looks like a unified architecture, so my original post doesn't fit so well. I ought to look at the patent proper next time!

XBD : The time taken to pass through the patent system is fairly irrelevant as it's the application date from which the patent is valid (if granted). Hence you can go ahead developing an idea and if someone copies it without a patent having been granted, once the patent is granted you sue the pants of 'em!

 

[mckmas8808]

So, since everybody here agrees that this is an unified shader GPU; what GPU do you think this maybe used for?

Any guesses?

 

[Carl B]

XBD : The time taken to pass through the patent system is fairly irrelevant as it's the application date from which the patent is valid (if granted). Hence you can go ahead developing an idea and if someone copies it without a patent having been granted, once the patent is granted you sue the pants of 'em!

Oh I know, the idea was 'fully developed' during the .com era but our company was forced to go into suspended animation due to a freeze on funding. If and when it clears is the determiner for what happens next on that front. I'm living my life in the meantime of course, but I mean, at the same time I'd like some movement on that front.

I think we'd sell it to a holding company we'd set up at this point and license from there, rather than try to go back into the mix. We've all kind of drifted into different fields now professionally, so seems like the best course of action to monetize on it. Four years from now of course... if then.

 

[Titanio]

So, since everybody here agrees that this is an unified shader GPU; what GPU do you think this maybe used for?

Any guesses?

As already mentioned, G80 would seem to be the obvious candidate. Perhaps this thread should be bounced back to 3D Tech

 

[trinibwoy]

Sounds to me like G70 pipes where the texture units are tied in with the shader units, so you either texture or shade but not both, the idea being that if you're not texturing, a specialist texture-unit is silicon sitting idle.

Doesn't look like it. Of all these functions outlined in the patent - LOD calculation, sampling, address calculation and filtering, what is the G70 ALU capable of? It looks like the flexibility outlined in this patent is geared toward allowing for different Texture Unit configurations for different levels of hardware.

Also, it's doubtful that G70 has such flexibile data routing capabilities between ALU's (PCU's?) and the fixed-function texturing hardware.

 

[Carl B]

Well... I might as well start off with some questions here to stimulate things a bit.

1) How certain can we be that this is G80?

2) Is G80 NV50? Or was NV50 cancelled? (I forget which) Seeing the filing date on this patent, if NV50 was cancelled, then G80 may differ in some ways from the core concept described here within the patent.

3) Assuming that this patent describes G80 to a degree, is there any good reason why RSX would *not* incorporate any of this technology? What are the pros for and against? With an expected launch date of Summer '06, what would the expectations be on G80's tape-out at this point?

 

[j^aws]

Well... I might as well start off with some questions here to stimulate things a bit.

1) How certain can we be that this is G80?

Well, if you think G80 is unified then this would be applicable. I'm not sure if anyone has any concrete source that it is unified. I think earlier in this thread, Uttar was suggesting that G80 isn't?

2) Is G80 NV50? Or was NV50 cancelled? (I forget which) Seeing the filing date on this patent, if NV50 was cancelled, then G80 may differ in some ways from the core concept described here within the patent.

IIRC, it was only an internal name designation change. The cancellation of NV50 was a rumor, but there could well be more to it...

3) Assuming that this patent describes G80 to a degree, is there any good reason why RSX would *not* incorporate any of this technology? What are the pros for and against? With an expected launch date of Summer '06, what would the expectations be on G80's tape-out at this point?

Well judging by the file date , Jan '04... this tech may well have been on the table for Sony but everything indicates G7x, unless it's all a red herring!

Interestingly, the patent also describes a GPU with programmable ROPs, i.e. none!

...When Texture Unit 225 includes Load/Store Unit 840, execution pipelines within a Programmable Graphics Processing Pipeline 950 may be configured to perform near and far plane clipping and raster operations, such as stencil, z test, and the like. Results or samples output by Programmable Graphics Processing Pipeline 950 are saved in an output buffer stored in Local Memory 140 via Load/Store Unit 840. When the data received by Graphics Subsystem 170 has been completely processed by Graphics Processor 905, Output Controller 180 reads the output buffer to provide an Output 185 of Graphics Subsystem 170. Therefore, Raster Operation Unit 160 may be omitted from Graphics Processor 105. For example, a program may include program instructions for performing depth testing, including writing pixel data to the frame buffer via Load/Store Unit 840.

IIRC, I remember a Jen-Hsen interview referring to more programmable ROPs and expecting AA to be done in shaders etc. Well, I'm not sure about removing ROPs completely but I expect ROPs to be modified in RSX for reading/ writing to XDR/GDDR etc.

 

[Jawed]

http://www.beyond3d.com/forum/showpost.php?p=675926&postcount=11

Jawed

 

[Carl B]

IIRC, it was only an internal name designation change. The cancellation of NV50 was a rumor, but there could well be more to it...

I know that NV47/NV48 went to G70, but I wasnt sure if the indications were that NVidia 'changed directions' with NV50 (ie dropped it) and G80 is actually now the would-be NV60, or whether G80 still represents NV50. But I may be a little out of it here.

 

[RavenFox]

I think your correct xbdestroya. NV-50 was dropped for the jump to 60. Well from what I remembered anyway.
Some stuff from 50 could have been combined with 60. Thats as good a guess as any.

 

[Shifty Geezer]

Might it not also be worth considering that this patent doesn't cover a particular GPU either? Let's say it was something they were playing with when investigating unified shaders, and as is the norm patented one or more techniques on the way, but found the performance didn't match their discrete shader architecture. It could be an idea that isn't destined for any immediate GPU until they can solve the scheduling problems that Jen-Hsun has mentioned regards using US.

 

[overclocked]

I know that NV47/NV48 went to G70, but I wasnt sure if the indications were that NVidia 'changed directions' with NV50 (ie dropped it) and G80 is actually now the would-be NV60, or whether G80 still represents NV50. But I may be a little out of it here.

The right wording would be to simply call G70 NV45 IMO. Dont know why they called PCI-E version of NV40/45. Just messing up the cool codenames..

 

[j^aws]

http://www.beyond3d.com/forum/showpost.php?p=675926&postcount=11

Jawed

Just to be clear, he's referring to OM as ROPs right? It seems to me that programmable OM/ROPs are beyond DX10...

...It could be an idea that isn't destined for any immediate GPU until they can solve the scheduling problems that Jen-Hsun has mentioned regards using US.

Did Jen-Hsun specifically mention scheduling issues? If so, do you have a link?

 

[Jawed]

Just to be clear, he's referring to OM as ROPs right?

Effectively, as far as I can tell.

It seems to me that programmable OM/ROPs are beyond DX10...

And to Jack, which is why it's such a wistful thought Did you look at the thread as a whole?

There's a thread on the topic of shader AA:

http://www.beyond3d.com/forum/showthread.php?t=22472

including a link to David's comments.

Jawed

EDIT: whoops misconstrued Jen/David

 

[Shifty Geezer]

It was in a talk about why nVidia weren't interested in unified shaders, and IIRC it was the difficulties of 'scheduling' (not sure that's the right term) that had nVidia saying the time wasn't right for US at the moment.

Unfortunately I'm lousy at remembering when and where so can't provide a quote, but I'm sure someone else will know all the particulars!

 

[joebloggs]

[Version impression]

[/ Version impression]

 

[Carl B]

The right wording would be to simply call G70 NV45 IMO. Dont know why they called PCI-E version of NV40/45. Just messing up the cool codenames..

There already is an NV45 though: NV45

G70 is the new moniker for NV47.

EDIT: Ok I think I see what you're saying with NV45 as NV40, but when you remake the chip like that in order to be PCI-e native, I think it's deserving of a new chip designation; not to mention they fixed the PureVideo on that chip.