Semi Accurate's 4XX views

[trinibwoy]

I agree we will never see evidence of most architectural features, but that's no reason to blindly take what they say as fact. This is probably a dead end discussion, but what doesn't yet make sense to me is why you'd have 16 fixed function units when 4 would have been sufficient. Maybe it comes down to ease of implementation and the tessellator is small. Maybe it's marketing not telling us the real number of fixed function units. Who knows.

Yes, I agree it's a dead-end because there's no reason to feel the way you do. It's been pointed out explicitly in their slides and also discussed here at B3D that it makes sense to keep geometry units including the tessellators close to the shader core due to locality of data.

How do you know that 4 would have been sufficient without knowing anything about each unit's cost or throughput? Maybe each of these 16 has only 1/4 of the throughput of the 4 that you have in mind. The concerns seem really arbitrary. And you keep saying that marketing is lying but what do they gain by doing so and what have you seen that leads you to that conclusion? Maybe marketing just knows a lot more about their product than we do and all the consipracy theories are baseless?

 

[Picao84]

Yes, I agree it's a dead-end because there's no reason to feel the way you do. It's been pointed out explicitly in their slides and also discussed here at B3D that it makes sense to keep geometry units including the tessellators close to the shader core due to locality of data.

How do you know that 4 would have been sufficient without knowing anything about each unit's cost or throughput? Maybe each of these 16 has only 1/4 of the throughput of the 4 that you have in mind. The concerns seem really arbitrary. And you keep saying that marketing is lying but what do they gain by doing so and what have you seen that leads you to that conclusion? Maybe marketing just knows a lot more about their product than we do and all the consipracy theories are baseless?

Its quite clear in my mind that the reason for that discussion is: Charlies says so. If Charlie didnt say that everything was done in software, via the CUDA cores, i bet noone would be here questioning it.

 

[Malo]

Its quite clear in my mind that the reason for that discussion is: Charlies says so. If Charlie didnt say that everything was done in software, via the CUDA cores, i bet noone would be here questioning it.

Well to be fair, Rys mentioned something along the same lines when outlining the Fermi specs originally.

 

[PSU-failure]

In other words if the facts turned out to match his speculation then he would have been right. Brilliant, even when he's completely wrong it's not his fault, it's only because the facts didn't co-operate.

I didn't say "he was right", but "overall correct", though I made the exact same predictions without anything more than the previous generations as sources... and ended up overall correct and sometimes wrong, so he could very well have had next to zero "inside information", just some luck and mostly restospective analysis.

 

[Sontin]

I dont remember that. If he did, why don't you produce a link to it?

http://forum.beyond3d.com/showpost.php?p=1406715&postcount=65

http://www.theinquirer.net/inquirer/news/1137331/a-look-nvidia-gt300-architecture

Offtopic: Guess, who is not out from Apple? Right

 

[willardjuice]

And we're done. There's already another thread dedicated to this crap.