Inside Fermi's graphics architecture

[Wreckage]

http://www.techreport.com/articles.x/17815

 

[tviceman]

That was probably the most comprehensive article I've read yet it's initial introduction at the end of September.

 

[faxon]

that was really quite a good article to read through. i already saw a lot of that info browsing around on the web but the back story behind the company was interesting as well lol

 

[Keysplayr]

I keep noticing that Rys insists that tesselation is being done, or going to be done through software. In one of the weekly Nvidia questions, this question was answered by Jason Paul. Tesselation has dedicated hardware. Even says, "Sorry Rys "
Mebbe I should shoot Rys an email with a link to the weekly questions.
The rest of the article was a decent read though.

 

[yh125d]

bookmarked to read tomorrow at work

 

[shangshang]

oh shoot Wreckage is on my ignore. actually something was wrong with the forum when I didn't see any article! lol

 

[SirPauly]

I keep noticing that Rys insists that tesselation is being done, or going to be done through software. In one of the weekly Nvidia questions, this question was answered by Jason Paul. Tesselation has dedicated hardware. Even says, "Sorry Rys "
Mebbe I should shoot Rys an email with a link to the weekly questions.
The rest of the article was a decent read though.

Rys responded here with more clarity on this:

http://www.rage3d.com/board/showpost.php?p=1336090006&postcount=948

Since I wasn't terribly clear about it in the article, let me be clear about what I mean by it being a software pipe. I simply mean that there's no dedicated fixed block for the entirety of the D3D11 tessellation pipe in hardware, and that's the case for ATI too.

The comparison comes in the amount of fixed logic dedicated to tessellation computation and the data flow. I believe NV uses less logic here than ATI comparatively speaking, but both architecture run a 'software' pipe nonetheless.

Tell me that VS+HS and DS aren't prime candidates for threads running almost entirely on an ATI SIMD. Please! It's only TS that should get any significant help from fixed logic.

 

[Keysplayr]

Quote:
Originally Posted by Keysplayr
I keep noticing that Rys insists that tesselation is being done, or going to be done through software. In one of the weekly Nvidia questions, this question was answered by Jason Paul. Tesselation has dedicated hardware. Even says, "Sorry Rys "
Mebbe I should shoot Rys an email with a link to the weekly questions.
The rest of the article was a decent read though.

Rys responded here with more clarity on this:

http://www.rage3d.com/board/showpost...&postcount=948



Quote:
Originally Posted by Rys
Since I wasn't terribly clear about it in the article, let me be clear about what I mean by it being a software pipe. I simply mean that there's no dedicated fixed block for the entirety of the D3D11 tessellation pipe in hardware, and that's the case for ATI too.

The comparison comes in the amount of fixed logic dedicated to tessellation computation and the data flow. I believe NV uses less logic here than ATI comparatively speaking, but both architecture run a 'software' pipe nonetheless.

Tell me that VS+HS and DS aren't prime candidates for threads running almost entirely on an ATI SIMD. Please! It's only TS that should get any significant help from fixed logic.

Bold above: How can Rys possibly know that there is, or isn't a dedicated "fixed block" for the D3D11 "tesselation pipe" If he hasn't seen hide nor hair of Fermi? How can one make a call like that unless seeing, in unimaginable detail, every aspect of Fermi's architecture? He doesn't even know the rough specs entirely, let alone a super detail like this. I can understand him knowing about RV870, it's out, no secrets there.

 

[v8envy]

I think the hair being split is whether there is fixed function hardware just for tesselation, end to end. When people think software they immediately assume code running on the CPU. If there is some fixed function hardware to take tesselation API calls and turn them into shader code that would simultaneously be a "software" implementation and a "dedicated hardware" implementation, depending on which point you're trying to prove.

How he knows which approach was taken is a good question. I don't think you'd need "unimaginable detail" view to determine if there's a bunch of silicon dedicated to a particular fixed function or not, though.

 

[Keysplayr]

I think the hair being split is whether there is fixed function hardware just for tesselation, end to end. When people think software they immediately assume code running on the CPU. If there is some fixed function hardware to take tesselation API calls and turn them into shader code that would simultaneously be a "software" implementation and a "dedicated hardware" implementation, depending on which point you're trying to prove.

How he knows which approach was taken is a good question. I don't think you'd need "unimaginable detail" view to determine if there's a bunch of silicon dedicated to a particular fixed function or not, though.

Ok, but I don't think any dude can just look at a die shot of a GPU and say exactly what "that particular block" of transistors is supposed to do without a little guidance from the dudes who actually created the GPU.

In the end, we actually wouldn't truly know which way would be better or worse anyway. Dedicated full hardware, or via software. I dunno.

 

[Kuzi]

Wrong thread

 

[Idontcare]

Ok, but I don't think any dude can just look at a die shot of a GPU and say exactly what "that particular block" of transistors is supposed to do without a little guidance from the dudes who actually created the GPU.

In the end, we actually wouldn't truly know which way would be better or worse anyway. Dedicated full hardware, or via software. I dunno.

Is the term "via software" here used in reference to distinguish between "fixed hardware" and "programmable hardware"?

 

[Keysplayr]

Is the term "via software" here used in reference to distinguish between "fixed hardware" and "programmable hardware"?

I would say, yes. Fixed hardware being a dedicated tesselation "unit", and progrmmable hardware being say "shaders".