• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

AMD Polaris architecture – GCN 4.0

Ayahuasca said:
Anyone who thought it was x10 gaming performance was trolling or seriously uneducated in this sector.
Click to expand...

Yeah, I am sure if Nvidia did find a way of upping the speed Of Maxwell by 10 for their next card.....well it wouldn't be their next card. They would find a way to drip feed the power over to us in something like 3 year cycles or thereabouts to maximise profits. Same with AMD.

:D
 
It was said that its 10X Maxwell in Mixed Precision. Not that its a big thing, it is because Maxwell is weak in this.

This was the original slide:
NVIDIA-Pascal-and-Volta-Compute-Performance-635x357.jpg


I've edited Maxwell in it:
33013_dualpreckicsi.jpg


:)
 
Mtom said:
One more piece of information: Vega is a 2 GPU family just like Polaris.
Click to expand...

Since when? the only known vega part is greenland/vega10. Like they could release another part under the vega name after, but the part that will end up launching in october if the rumours are true are the greenland based parts. You only have to look at software support to see that only a single vega part will be launching anytime soon.
 
Last edited:
Those figures don't seem right to me, especially for the apparent die sizes being bigger than a direct shrink of Hawaii and Pitcairn.
 
Mauller said:
Those figures don't seem right to me, especially for the apparent die sizes being bigger than a direct shrink of Hawaii and Pitcairn.
Click to expand...

At least they ain't banging on about 850Mhz limits now :D

Pretty good base clocks, if true my guess is the new architectures is getting more from less stream processors.
 
I just had a thought, was that Polaris 11 matching a GTX 950 @ 850Mhz?

1400Mhz would be +65%, what does that equat to?
 
Mtom said:
It was said that its 10X Maxwell in Mixed Precision. Not that its a big thing, it is because Maxwell is weak in this.


:)
Click to expand...

Its not actually the double precision where a lot fo the performance came form, although that is there. It is actually the ability to have half-precision 16bit FP, which for some computing tasks, especially network weights and activation values in neural networks offer plenty of resolution. FP16 it directly twice as fast as FP32 and uses half the memory, which allows better use of ache etc.

Another piece of the story is simply moving to HBM2 and the large increase in bandwidth. For compute applications memory bandwidth is much more important because different data sets are consonantly being addressed, cf. a game where although a level might have a lot of total asset the amount of memory used 1 frame to the next is smaller.


NVlink was another piece of the performance puzzle, allowing faster scaling in multi-gpu setups.

Then there is the usual 2x speed up for a new process, a lot of other architectural changes.


And then there is the fact that Nvidia didn't claim it in a serious manner, only referring to all these individual 2x-4x speed improvements makes Pascal seem lie 10x Maxwell for compute. The actually numbers are provided with benchmarks from Alexnet etc.
 
You must log in or register to reply here.
Back
Top Bottom