** The Official Nvidia GeForce 'Pascal' Thread - for general gossip and discussions **

[Jono8]

Ground breaking is a bit strong word...bringing advancement in nothing but power efficiency...in performance they are can't really overcome the much older amd cards.

What? The 980Ti is faster than the Fury X whilst using less power WITHOUT hbm. It is pretty impressive architecture. AMD had to jump early into hbm to get even close to maxwell in terms of perf per watt.

Im not saying fiji isnt good, its just maxwell is impressive because it at the very least matches ,if not beats fiji in perf and perf per watt even without hbm.

 

[Harlequin]

Nv need to do something as DX12 performance is lacking badly when a 290x is matching the latest and best. Gp100 is going GCN is obvious from the fore layout

 

[Mtom]

What? The 980Ti is faster than the Fury X whilst using less power WITHOUT hbm. It is pretty impressive architecture. AMD had to jump early into hbm to get even close to maxwell in terms of perf per watt.

Im not saying fiji isnt good, its just maxwell is impressive because it at the very least matches ,if not beats fiji in perf and perf per watt even without hbm.

The world is not just 980Ti and Fiji.
I said as well it is an improvement in power efficiency, but not a "ground breaking" architecture in my book.
980 is barely faster than a much older GCN (390X), the 970 is slower than a much older GCN (390), the 960 is slower than an older GCN (380X)
Especially not ground breaking in DX12 where in most titles released so far the 970 is not battling the 390 instead it is nosediving to R9 380 territory, and the 980 is left behind by the 390 and 390X.

 

[humbug]

What? The 980Ti is faster than the Fury X whilst using less power WITHOUT hbm. It is pretty impressive architecture. AMD had to jump early into hbm to get even close to maxwell in terms of perf per watt.

Im not saying fiji isnt good, its just maxwell is impressive because it at the very least matches ,if not beats fiji in perf and perf per watt even without hbm.

As Harlequin above also pointed out that is subjective, even in DX11 at 4K Fury-X often surpasses 980TI performance, so one could argue the Fury-X is the better card.
I'm not suggesting that as i do agree in many ways Maxwell is better, given that the reason for the 980TI better performance at 1080P is due to Maxwell's better instruction scheduling.
There are aspects of the Fiji architecture that clearly more advanced than Maxwell, Fiji, even Hawaii's parallel compute engine is far more sophisticated than anything Nvidia have, and may be even with Pascal.

One is not better than the other cuz (insert specific aspect here while ignoring others)
They are just different. better or worse in different ways.

As for power consumption, not really, the Fury-Nano manages to be faster than the 980 while consuming the same level of power, does that mean GCN is more efficient than Maxwell? no.

 

[Mtom]

As Harlequin above also pointed out that is subjective, even in DX11 at 4K Fury-X often surpasses 980TI performance, so one could argue the Fury-X is the better card.
I'm not suggesting that as i do agree in many ways Maxwell is better, given that the reason for the 980TI better performance at 1080P is due to Maxwell's better instruction scheduling.
There are aspects of the Fiji architecture that clearly more advanced than Maxwell, Fiji, even Hawaii's parallel compute engine is far more sophisticated than anything Nvidia have, and may be even with Pascal.

One is not better than the other cuz (insert specific aspect here while ignoring others)
They are just different. better or worse in different ways.

As for power consumption, not really, the Fury-Nano manages to be faster than the 980 while consuming the same level of power, does that mean GCN is more efficient than Maxwell? no.

Agree with that...Maxwell is great in DX11, and great in perf/W, but it lacks a lot in DX12 performance. GCN is the opposite, less power efficient, but it is because it ready to take on DX12

 

[ICDP]

Nv need to do something as DX12 performance is lacking badly when a 290x is matching the latest and best. Gp100 is going GCN is obvious from the fore layout

Though according to rumoured info GP100 is NOT consumer Pascal.

 

[Harlequin]

Though according to rumoured info GP100 is NOT consumer Pascal.

rumour mill says consumer pascal is a die shrink Maxwell core.....

 

[Jono8]

As Harlequin above also pointed out that is subjective, even in DX11 at 4K Fury-X often surpasses 980TI performance, so one could argue the Fury-X is the better card.
I'm not suggesting that as i do agree in many ways Maxwell is better, given that the reason for the 980TI better performance at 1080P is due to Maxwell's better instruction scheduling.
There are aspects of the Fiji architecture that clearly more advanced than Maxwell, Fiji, even Hawaii's parallel compute engine is far more sophisticated than anything Nvidia have, and may be even with Pascal.

One is not better than the other cuz (insert specific aspect here while ignoring others)
They are just different. better or worse in different ways.

As for power consumption, not really, the Fury-Nano manages to be faster than the 980 while consuming the same level of power, does that mean GCN is more efficient than Maxwell? no.

You missed my point which was that maxwell does all that without all the benefits of hbm. It is a very impressive architecture in those terms. Imagine maxwell with the bandwith and power consumption benefits of hbm. Unfortunately we didnt get to see that as they new they didnt have to with what the competition was offering.

Hopefully AMD will and have upped their game a bit for this next generation. Nvidia have had it too easy on 28nm.

 

[Harlequin]

You missed my point which was that maxwell does all that without all the benefits of hbm. It is a very impressive architecture in those terms. Imagine maxwell with the bandwith and power consumption benefits of hbm. Unfortunately we didnt get to see that as they new they didnt have to with what the competition was offering.

Maxwell sucks at DX12 - async which cannot be fixed in software - where are the windows 10 DX12 drivers for Fermi? it was promised to be a DX12 card yet they don't exost! same with Maxwell - it might be able to `do` some DX12 - butcast your mind back to the R100 radeon and GF3; 1 could do Pixel shaders and the other couldn't.

and which went on the lead the DX8 forefront.

 

[Rroff]

rumour mill says consumer pascal is a die shrink Maxwell core.....

Same "source" also claimed not so long ago that GP100 had 6144 CUDA cores so not sure I'd take it too seriously.

Maxwell sucks at DX12 - async which cannot be fixed in software - where are the windows 10 DX12 drivers for Fermi? it was promised to be a DX12 card yet they don't exost! same with Maxwell - it might be able to `do` some DX12 - butcast your mind back to the R100 radeon and GF3; 1 could do Pixel shaders and the other couldn't.

and which went on the lead the DX8 forefront.

I'm not sure what nVidia are doing with Maxwell as while there are some issues with async (which largely absolutely can be solved in software aslong as you can avoid penalties from some additional degree of context switching required) the only real issue is that it takes more developer hands on than the AMD solution to get the same result. Possibly nVidia are stringing AMD along or holding back to try and force people to move to Pascal when it arrives :S

Fermi is largely screwed for DX12 anyhow due to the lack of any hardware ability to execute more than 1 type of compute, graphics or memory op simultaneously (which will be a bigger deal when developers start to use DX12 proper rather than just wrapping their current engines).

 

[Harlequin]

http://nvidia.custhelp.com/app/answers/detail/a_id/3711/~/which-nvidia-gpus-will-support-dx12?


don't think it will ever happen

don't get me wrong - after having 2 xgtx980 in sli ( and now 1 since it took 6 months to fix sli - btw Diplay Port is still broken as it drops back to 30hz in some content @4k)

GTX 980 is a great DX11 card - but it gains nothing from DX12 , and if you start adding software scheduling into the mix there will be a penalty ; you can only throw so much dev time at an issue before you have to say ` this wont work`. look at gameworks in batman for example - the dev has given up with it on pc as its still broken.

 

[humbug]

You missed my point which was that maxwell does all that without all the benefits of hbm. It is a very impressive architecture in those terms. Imagine maxwell with the bandwith and power consumption benefits of hbm. Unfortunately we didnt get to see that as they new they didnt have to with what the competition was offering.

Hopefully AMD will and have upped their game a bit for this next generation. Nvidia have had it too easy on 28nm.

In most cases i don't think HBM1 makes all that much difference vs fast GDDR5 TBH.

Memory Bandwidth:
Fury-X: 512GB/s
980TI: 336.5GB/s

That extra bandwidth isn't going to help unless its needed, it wouldn't be needed at 1080P, or so much at 1440P.

The power consumption of HBM vs good Samsung Memory IC's is also likely to be pretty negligible.

HBM is only going to come into its own once GPU's get powerful enough to run extremely high resolutions at high frame rates.

I think AMD will look at where they are lacking, in fact Raja Koduri did say they have redesigned their own Instruction Scheduling and Tessellation handling, essentially using the same system Nvidia are in Maxwell.

 

[bru]

The world is not just 980Ti and Fiji.

+1

980 is barely faster than a much older GCN (390X), the 970 is slower than a much older GCN (390), the 960 is slower than an older GCN (380X)

Most of us will know what you mean, but that is such a bad way to put it. You have listed cards 980, 970 (released 18.09.14) and 960 (22.01.2015) which are all Maxwell V2 which made its first appearance on 18.09.14 and compared them against cards 390 390X (released on 18.06.2015) and 380X (released on 19.11.2015) which are GCN 1.1 and GCN 1.2 which first made their appearances on 22.03.13 with the 7790 and 02.09.14 with the 285.

I would suggest either comparing by architecture or by card, not trying to mismatch both.

 

[Dave2150]

Ground breaking is a bit strong word...bringing advancement in nothing but power efficiency...in performance they are can't really overcome the much older amd cards.

Maxwell destroys Hawaii and Fiji at 1080P and 1440P in DX11 titles (99.9% of games). It's only at 4k that Fiji is competitive with Maxwell.

DX12 is a different story, though by the time any significant amount of titles are released, Polaris and Pascal will be here, making this debate pointless.

 

[Aretak]

Maxwell destroys Hawaii and Fiji at 1080P and 1440P in DX11 titles (99.9% of games).

That's nonsense, frankly. The 390 beats out the 970 in most DX11 titles these days, even at 1080p, and the 380 and 380X are usually comfortably ahead of the 960. The Nano is also generally at least level with the 980. Across the two product stacks it's pretty even in terms of the number of cards outperforming each other at different price points.

 

[Perfect_Chaos]

That's nonsense, frankly. The 390 beats out the 970 in most DX11 titles these days, even at 1080p, and the 380 and 380X are usually comfortably ahead of the 960. The Nano is also generally at least level with the 980. Across the two product stacks it's pretty even in terms of the number of cards outperforming each other at different price points.

He's talking about the 2 high end cards but didn't mention it in that post, i guess.

 

[andybird123]

You're comparing a 366mm2 chip (the 380/X)
with a 227mm2 chip (the 960)

Obviously as a consumer you are interested in price first and foremost, but from a business/profit point of view, that 960 is wiping the floor with the 380/X

 

[egraphixstudios]

"fight fight"....

 

[ICDP]

He's talking about the 2 high end cards but didn't mention it in that post, i guess.

He specifically mentioned architectures not individual models of GPU. Maxwell is competing against Hawaii and Fiji and across the price stack (apart from 980Ti vs Fury X) AMD are actually comparing very well to their Nvidia counterparts in DX11 and generally exceeding them in DX12.

 

[andybird123]

consumer price or cost to produce ^ Hawaii and Fiji are needing much larger die sizes to compete on performance, Maxwell is an incredibly efficient design by comparison

at 40-60% larger die sizes they very well should be beating the "equivalent" Maxwell cards