Pascal architecture question

I was looking at the configurations of the various laptop Pascal GPUs (summarized here on Wikipedia) and I was wondering what impact the different ratios of components have. Here's a summary table:

Card : Shaders : Texture Mapping Units (TMUs) : Render output units (ROPs)
1080: 2560:160:64 (80:5:2)
1070: 2048:128:64 (64:4:2)
1060: 1280:80:48 (80:5:3)
1050Ti: 768:48:32 (48:3:2)
1050: 640:40:16 (80:5:2)

The numbers in parenthesis are reduced ratios to make it easier to compare. As you can see, there are always 16 shaders per TMU, but the number of ROPs varies and not in a predictable way: there are 40 shaders per ROP on the 1080 and this number decreases as we go down the list... but then the 1050 has 40 shaders per ROP again.

So, why are they configured this way and which ratio is optimal? In particular, the 1050 is effectively the 1080 divided by 4 across the board while the 1060 is the 1080 divided by 2... but there are 1.5 times as many ROPs as one would expect.

 

Quick question, do you know what an "SMX" is?

 

ROP partitions to memory controllers is 1:1. That's why 1060 (192-bit) has 3/4 as many ROPs as 1070/1080 (256-bit), 1050/1050 Ti (128-bit) has 1/2 as many, and Titan XP (384-bit) has 1-1/2 as many.

 

That explanation almost accounts for it... but then why does the 1050 have half the ROPs of the 1050 Ti when both of them have a 128-bit memory bus?

 

It has 32 ROPs just like 1050 Ti...

 

As far as I can tell, it's Nvidia's name for a streaming multiprocessor ever since Kepler.

That is really interesting. In that case, everything makes sense, but it would mean that Wikipedia, AnandTech and TechReport are all wrong about the laptop 1050 -- both of the linked articles explicitly point out that the laptop 1050 has 16 ROPs thus making it different from both the laptop 1050 Ti and the desktop 1050. I guess there must have been a mistake in the CES-era Nvidia info that got propagated everywhere. Or... how confident are you in that screenshot? It shows 53 for the number of TMUs which I'm pretty sure can't be right.

 

Interesting. I wasn't aware that Nvidia disabled half the ROPs on the 1050N.

That GPU-Z says 32 ROPs/40 TMUs and is from a DT 1050.

 

Thanks. Indeed, the desktop card has 32 ROPs in its specifications and it's good to see that it is the case. I'm going to go ask in the manufacturer forums for laptop 1050 screenshots -- if your explanation is correct, Nvidia has given it ROPs as if it had a 64-bit bus which would be kind of weird.

 

So, thanks to exile85, we finally have a valid screenshot of the laptop GTX 1050 (as found in the Dell XPS 15). As you can see, it has 32 ROPs just like its desktop counterpart and the 1050 Ti cards. The info about 16 ROPs is a typo from the CES material which was propagated all over the place.