AMD Polaris architecture – GCN 4.0

pmc25 · 22 Apr 2016 at 10:23

Mtom said:
From a marketing viewpoint. All P10 cards will have 8GB i think. Just look at the flak the Fury got because of it. On the smaller cards 4GB would do fine, but the price is not high enough to justify HBM.

I think almost everyone has forgotten capacity issues, it's proved to be a total non-issue. Only thing imo that might effect it is HDR, which AMD are going to push and NVIDIA presumably won't or don't support.

Fairly sure NV won't support that (or will try to fudge it in software) until Volta.

AMD brought forward support because Sony need it to sell their top of the range TVs.

But back to the point. Though Polaris seems to be less bandwidth hungry, a bigger chip would need more than GDDR5 can feasibly offer at 8GB. I doubt they return to a 512bit gddr bus, as likely that proves more expensive than hbm1 at this point. With the 28 to 14nm shrink, there should be enough room on an interposer smaller than Fiji's to do 6GB or 8GB.

Overlag · 22 Apr 2016 at 12:13

Diagro said:
I was under the impression that there wasn't a 4GB Cap....

with HBM1 there is, unless you want to increase the bus width. Each GB uses 1024bits, hence the 4096bit bus. HBM2 fixes that.

humbug · 22 Apr 2016 at 12:16

Interesting snipit in this article....

As for AMD's current businesses, sales were down 19% Y/Y in Q1, thanks to PC weakness, share loss to Intel and Nvidia (NASDAQ:NVDA), and lower sales of console-related processors. But AMD expects Q2 sales to rise ~15% Q/Q, thanks largely to a new semi-custom SoC win (believed to be for a 4K-capable PS4) and higher GPU sales (aided by the Polaris GPU launch). Moreover, the company forecasts $1.5B in revenue from three different semi-custom SoC wins over the next 3-4 years. Following a brutal 2014 and 2015, there's finally some hope for a sales bottom.

So if AMD are expecting Polaris sales to be represented in Q2 earnings they must go on sale in Jun.

http://seekingalpha.com/article/396...1c2:932392da5d231925ea2ca29a9870ae01&uprof=55

Jono8 · 22 Apr 2016 at 12:34

I too hope for a sales bottom.

CAT-THE-FIFTH · 22 Apr 2016 at 16:35

I was reading the latest financial results this quarter and it appears AMD had double digit increases in GPU sales. Lets hope AMD can build on it this year and try to get back to 30% to 35% of volume of GPUs shipped in the next 12 months.

CAT-THE-FIFTH · 22 Apr 2016 at 16:50

Just saw the following comment on AT forums.

zlatan said:
There are some situation where Polaris is incredibly fast. Faster than anything in the market. The secret is probably that special culling mechanism in the hardware, which helps the GPU to effectively cull those false positive primitives that aren't visible in the screen. Today's hardwares can't do this.
Single wavefront perfomance is also incredibly good. 10-100 times faster than anything in the market. This is good for VR.

The poster in question has been unusually accurate about certain things over the last few years(I believe they even were the first person to say AMD would be making a 500+ MM2 GPU),so interesting words.

Hinted Polaris 10 would be quicker than a Fury "depending on the applications".

So Polaris 10 around R9 390 to Fury level performance but with potential to be faster than Fury with newer games,ie,DX12??

queamin · 22 Apr 2016 at 17:14

CAT-THE-FIFTH said:
Just saw the following comment on AT forums.

The poster in question has been unusually accurate about certain things over the last few years(I believe they even were the first person to say AMD would be making a 500+ MM2 GPU),so interesting words.

Hinted Polaris 10 would be quicker than a Fury "depending on the applications".

So Polaris 10 around R9 390 to Fury level performance but with potential to be faster than Fury with newer games,ie,DX12??

By all counts gcn4 has a lot of new and improvement over this gen as a lot of it has been updated.

LoadsaMoney · 22 Apr 2016 at 18:37

AMD Polaris To Take Share From Nvidia – Launch Confirmed In Q2 2016

http://wccftech.com/amd-graphics-share-rival-nvidia-2016-report/

Rroff · 22 Apr 2016 at 19:11

CAT-THE-FIFTH said:
Just saw the following comment on AT forums.

zlatan said:

The secret is probably that special culling mechanism in the hardware, which helps the GPU to effectively cull those false positive primitives that aren't visible in the screen. Today's hardwares can't do this.

Click to expand...

Eh what? the problem with hardware occlusion culling has always been the balancing act between hitting drawcall limitation/stalling the CPU and/or pulling down performance through that versus the gains from culling - you can't go indiscriminately culling primitives or pixels (samples) in hardware without knowing what the programmer intends - something that low level APIs would help to utilise better.

CAT-THE-FIFTH · 22 Apr 2016 at 19:12

Rroff said:
Eh what? the problem with hardware occlusion culling has always been the balancing act between hitting drawcall limitation/stalling the CPU and/or pulling down performance through that versus the gains from culling - you can't go indiscriminately culling primitives or pixels (samples) in hardware without knowing what the programmer intends - something that low level APIs would help to utilise better.

Supposedly Nvidia has already been doing this for a while??

Rroff · 22 Apr 2016 at 19:17

CAT-THE-FIFTH said:
Supposedly Nvidia has already been doing this for a while??

nVidia has fairly efficient hardware culling - not really sure on whether there is potential for more efficient hardware culling of primitives compared to samples which is already fairly efficient but the problem has always been the software interface rather than the efficiency of the hardware implementation.

hominid · 22 Apr 2016 at 19:24

You can by depth of field, there is no point drawing something that is either fully of partially hidden by something else in front of it - witch is what happens atm

Rroff · 22 Apr 2016 at 19:28

hominid said:
You can by depth of field, there is no point drawing something that is either fully of partially hidden by something else in front of it - witch is what happens atm

That assumes the programmer isn't doing something outside the box.

CAT-THE-FIFTH · 22 Apr 2016 at 19:40

Rroff said:
nVidia has fairly efficient hardware culling - not really sure on whether there is potential for more efficient hardware culling of primitives compared to samples which is already fairly efficient but the problem has always been the software interface rather than the efficiency of the hardware implementation.

From what I gather AMD can't do this in hardware with its current GPU so it is something new for an AMD GPU. I assume he was referring to current AMD GPUs which don't do it.

Rroff · 22 Apr 2016 at 19:46

CAT-THE-FIFTH said:
From what I gather AMD can't do this in hardware with its current GPU so it is something new for an AMD GPU. I assume he was referring to current AMD GPUs which don't do it.

A lot of games use techniques like PVS to cull geometry anyhow - though they aren't massively efficient due to being run largely on the CPU and/or a certain amount pre-computed for a game level but it does reduce the potential gains from hardware culling of primitives quite a lot - for instance though you'd still often have the geometry of a player character and other dynamic elements being processed even when they are hidden by the world geometry.

CAT-THE-FIFTH · 22 Apr 2016 at 19:57

Rroff said:
A lot of games use techniques like PVS to cull geometry anyhow - though they aren't massively efficient due to being run largely on the CPU and/or a certain amount pre-computed for a game level but it does reduce the potential gains from hardware culling of primitives quite a lot - for instance though you'd still often have the geometry of a player character and other dynamic elements being processed even when they are hidden by the world geometry.

OK,but OTH AMD cards do tend to be weaker than Nvidia when it comes to tessellation,so I wonder if this is more a reaction to things like Gameworks effects,many of which emphasis use of tessellation??

pmc25 · 23 Apr 2016 at 00:45

CAT-THE-FIFTH said:
Just saw the following comment on AT forums.

The poster in question has been unusually accurate about certain things over the last few years(I believe they even were the first person to say AMD would be making a 500+ MM2 GPU),so interesting words.

Hinted Polaris 10 would be quicker than a Fury "depending on the applications".

So Polaris 10 around R9 390 to Fury level performance but with potential to be faster than Fury with newer games,ie,DX12??

zlatan very rarely posts anything he doesn't have detailed knowledge of. if he said that, it's likely true. if i recall, he's a developer involved with VR ... if the company he works for is big enough, they probably have samples already.

Boomstick777 · 23 Apr 2016 at 05:30

AMD stock has doubled (http://www.nasdaq.com/symbol/amd) good to see them gaining traction.

Freddie1980 · 23 Apr 2016 at 10:31

52%...... someone must be laughing all the way to the bank!!!!! I wish i followed my own advice and bought some AMD stock late last year when they announced Zen.

muziqaz · 23 Apr 2016 at 10:35

Freddie1980 said:
52%...... someone must be laughing all the way to the bank!!!!! I wish i followed my own advice and bought some AMD stock late last year when they announced Zen.

yeah, same old scamming financial market