help me here with g71 v r580

[seeker010]

with the Inquirer's explanation http://www.theinquirer.net/?article=28609, I'm very confused as to what they mean. I thought nvidia also decided that fragment operations were more important and thus increased the number of pixel alus without increasing the ROPs.

 

[Skrying]

It says 16 ROPs, that would be the same as G70.

Also, doing a comparison right now is rather pointless, you cant do it based on rumors, though some might be clear, you just should avoid it.

 

[Fodder]

I'm very confused as to what they mean.

Don't worry, so is Fuad.

 

[neliz]

Don't worry, so is Fuad.

Lol

My opinion is that under general circumstances, the new cards WILL be faster than the old cards... and THG will show the X850 leading ALL the other cards in HalfLife2...

 

[seeker010]

right, but my point is the inquirer is stating that ati and nvidia are taking different approaches. It is my understanding that both are increasing pixel alus without increasing theoretical pixel output.

 

[Ailuros]

right, but my point is the inquirer is stating that ati and nvidia are taking different approaches. It is my understanding that both are increasing pixel alus without increasing theoretical pixel output.

Notify me when Fuad manages to keep ROPs from ALUs apart. Other than that both G7x@90nm and R580 are likely to end up with 16 ROPs; as for ALUs one scenario would be 64 ALUs for the first and 48 for the second.

 

[Fodder]

right, but my point is the inquirer is stating that ati and nvidia are taking different approaches.

Fuad is, as usual, saying an awful lot without actually saying anything. As Ailuros said, the article appears to suggest NVIDIA's ROPs are somehow the same as ATI's 'shaders', which is not the case. As I understand it, ATI have decoupled PS ALUs from TMUs, and so are ramping up the shader power significantly but leaving TMUs and ROPs alone. With NV, the TMUs and ALUs are still bolted together, so both will go forth and multiply, while they may be content to stick with 16 ROPs.

 

[compres]

Notify me when Fuad manages to keep ROPs from ALUs apart. Other than that both G7x@90nm and R580 are likely to end up with 16 ROPs; as for ALUs one scenario would be 64 ALUs for the first and 48 for the second.

I think the 48 number comes from shader units rather than alus, so your 64 > 48 number might be incorrect. It might be 64 < 96 if I am correct.

 

[Pete]

Ail is counting just MADDs, which I think he did before to simplify things. I don't know the frequency of instruction use, but it might be fairer to call R580 48*1.5 b/c it has the extra ADD on the mini-ALU. Then again, I also don't know how specialized hardware helps (FP16 NRM, etc.), or how NV's integrating the TMU on the first ALU hurts.

Then again (edit: again), I'm still not sure how flexible NV's dual ALU shader pipe is, so calling it x*2 might even be shortchanging it (edit: I'm thinking mainly of co-issue).

 

[ANova]

Yeah, G71 is 32 pixel processors with 2 ALUs per processor. R580 is 48 pixel processors with one full ALU and one mini ALU per processor. Both have 16 ROPs and TMUs, although there is talk that the G71 may have 20 or 24.

 

[Ailuros]

I think the 48 number comes from shader units rather than alus, so your 64 > 48 number might be incorrect. It might be 64 < 96 if I am correct.

It's 4 quads with 48 ALUs; no idea what makes you call them "shader units" and not an arithmetic logic unit. In any case there are articles here at B3D available that analyze the specifics of both G70 and R520 to take those as a possible reference for upcoming refreshes from either side, where all the details of ALU/texturing/shading processors are being layed out in detail.

 

[Ailuros]

Ail is counting just MADDs, which I think he did before to simplify things. I don't know the frequency of instruction use, but it might be fairer to call R580 48*1.5 b/c it has the extra ADD on the mini-ALU. Then again, I also don't know how specialized hardware helps (FP16 NRM, etc.), or how NV's integrating the TMU on the first ALU hurts.

Then again (edit: again), I'm still not sure how flexible NV's dual ALU shader pipe is, so calling it x*2 might even be shortchanging it (edit: I'm thinking mainly of co-issue).

I'm staying away from mini-ALU yadda yadda because it's truly complicating things too much for my taste:

Here the mini-ALUs have according to the B3D article added MULL functionalities.

The real essence of the article is rather in this paragraph:

....although NV40 has two ALU's they are not in fact each fully featured with the same instructions, instead one is a MADD unit and the other is a MUL unit; for G70 NVIDIA say they have added a MADD and MULL into each of the units that didn't previously contain them and in fact we are led to believe they are now complete instruction duplicates of each other (although, obviously the second unit doesn't have texture address processing instructions). The net result is that G70 features 48 fragment shaders of the same capabilities, with one of them having to handle the texture processing instructions. The pipelines are arranged as a single SIMD array and they will all be operating on data of the same state.

http://www.beyond3d.com/previews/nvidia/g70/index.php?p=02