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

[Mauller]

I really don't understand how people think that AMD can get to 980ti/Furyx performance with the midrange part but NVidia cannot. both sides will have not too dissimilar performance, as is the usual state of affairs.

In a relative sense regardless of node, Pascal will not clock as well as Maxwell does due to the addition of dark silicone. Now it more depends on how they implement Double precision as to how much of a hindrance this dark silicone is.

Nvidia in the past tended to use a dedicated DP unit for so many SP units where as AMD with GCN fuse two SP units to compute a DP instruction, hence why Hawaii and Fiji can perform 1/2 DP.

The thing with GCN1-3 is that is has more theoretical performance than Kepler and Maxwell parts of similar price bracket. GCN is just hindered in graphics workloads and shows its true performance when a decent amount of Async compute is also being used. Often making it match or surpass highly overclocked maxwell parts, while retaining stock clocks.

but with them adding many features and tweaking GCN with ver 4 in polaris, many of the issues will be reduced and the architecture should be able to make use of more of its cores in single threaded workloads as well as multi-threaded workloads.

So in other words, GCN4 parts with similar numbers of shaders will have far better performance than an equal shader and clocked GCN3 part. And this is all before any improvements in clock due to the node change.


What i am getting at is that it is far easier for AMD to find performance with Polaris parts than it will be for Nvidia and pascal. Unless they want to price down things like the 1080 to midrange then they are going to have issues if they don't have a better performing part.

 

[Jono8]

Sounds more realistic, but what's the source for this rumour?

No it doesn't! What is wrong with you people

Nvidia won't release a 1080 card (which will be their flagship for a while until full fat Pascal comes out) that isn't even as fast as a 980Ti. It would be utterly stupid.

This would mean that full fat Pascal would be barely even 30% faster than the 980Ti and likely release almost 2 years after.

That would make it the most disappointing generation ever especially as this is on new architecture AND on a die shrink.

 

[Jono8]

Their love of AMD prevents any kind.of logic. According to the logic of certain posters the 1080 will be hardly any faster than a 980 and almost as expensive as a 980ti. Ignore them.

It is bizarre. It is like logic and all evidence from the history of past gpu generations has been thrown in the bin.

 

[bru]

Pascal will not clock as well as Maxwell does due to the addition of dark silicone.

Ok that is a good one.

Of course, now we know why Polaris will be so bright it will burn your eyes out, it is the ground up ground up testies of fresh born male pixie children.

 

[CAT-THE-FIFTH]

Ok lets assume the new 1080 card is a little below a 980Ti in performance
and costs over 400, why would you not buy a 980Ti instead ( which should fall below £500 soon unless nvidia artificially inflates the price )
so unless people really care about power consumption these cards will need to be relatively cheap ( and if they did care about power draw they are still getting screwed because no HBM2 )

Gameworks.

 

[Mauller]

Ok that is a good one.

Of course, now we know why Polaris will be so bright it will burn your eyes out, it is the ground up ground up testies of fresh born male pixie children.

It is true, relatively Pascal will not clock as well since Maxwell was made to cater to gaming. Hence why it has such gimped DP performance and why nvidia does not quote any DP metrics for their quadro parts. Their only DP part is in their tesla cards which still run Kepler.

Cant make a logical comeback, act like a child and call them a fanboy. - Damage.Patrol (D.P) logic

 

[Perfect_Chaos]

So is the 1080Ti due later this year or next year? And the Fury x successor in the same time frame?

 

[Mauller]

So is the 1080Ti due later this year or next year? And the Fury x successor in the same time frame?

Pretty much next year for big parts. Still a new node for making large monolithic dies on. Even if they have some trickling out it wont be up to mass production levels yet for public sales.

 

[queamin]

I think it will be like the 970/980 and then 980ti/titanx, like Mauller said, I think Nvidia might get some large dies out early but not many.

 

[bru]

It is true, relatively Pascal will not clock as well since Maxwell was made to cater to gaming. Hence why it has such gimped DP performance and why nvidia does not quote any DP metrics for their quadro parts. Their only DP part is in their tesla cards which still run Kepler.

Cant make a logical comeback, act like a child and call them a fanboy. - Damage.Patrol (D.P) logic

LOL you started it with 'Dark Silicone'

You want a logical comeback.

Ok 'Pascal will not clock as well as Maxwell' I quite agree, but as to why this is the case for anyone is a guess.

If I had to guess, it would be because Pascal is the first (for NVidia) of a new process, whereas Maxwell is the last on a very mature process.
It may be that clocks speeds are actually higher, but the percentage that you can get overclocking will be less. I would even suspect that NVidia's boost function will leave less on the table this time around as well, but of course like everyone else this is just me guessing.

 

[bru]

Opps double post

 

[Mauller]

Yeah, any current GP100 parts will be in testing and/or will go into professional parts. Since Nvidia really needs to replace kepler in their tesla products and maxwell in their quadros.

 

[Mauller]

LOL you started it with 'Dark Silicone'

You want a logical comeback.

Ok 'Pascal will not clock as well as Maxwell' I quite agree, but as to why this is the case for anyone is a guess.

If I had to guess, it would be because Pascal is the first (for NVidia) of a new process, whereas Maxwell is the last on a very mature process.
It may be that clocks speeds are actually higher, but the percentage that you can get overclocking will be less. I would even suspect that NVidia's boost function will leave less on the table this time around as well, but of course like everyone else this is just me guessing.

In a gaming sense, DP units are Dark silicone since they are unused the majority of the time. And as i pointed out the limitations of their overclocking depends on how they implement DP and you only have to compare Kepler to Maxwell to see how DP can hinder overclocking when Dedicated DP units are used.

But even being dark silicone they still consume power when inactive as they are part of the cuda cores that are being used. And i doubt they had use of any kind of specific gating to cut power to the DP parts when not in use.

Although i agree that default clocks are more than likely going to be higher due to the node change.

 

[Rroff]

DP units are Dark silicone

*snigger*

 

[Ash0787]

yeah I think you guys are right, they will try and position the 1080 as the relpacement or upgrade to the GTX 970 since that is their money maker at the moment

It will likely have 6GB of VRAM and perform similar to 980Ti but not more, be slightly more expensive than current 970
and they will market it heavily to make it seem attractive despite its flaws
and somehow make the 980Ti less appealing

If they take this approach they create a danger that people will opt for cheap Fiji cards instead,
but as we know AMD does not intend to sell large number of Fiji, and AMD brand image is not strong at the moment

 

[Perfect_Chaos]

Pretty much next year for big parts. Still a new node for making large monolithic dies on. Even if they have some trickling out it wont be up to mass production levels yet for public sales.

Will be keeping hold of my 980Ti until then, at least it's a decent time to save up

 

[Mauller]

*snigger*

Please enlighten me in your infinite wisdom as to how dedicated DP units are used when a GPU performs graphics processing? Because in the majority they aren't and remain inactive but still using power. Hence Dark Silicone as it is silicone that is not being used.

 

[james.miller]

That is not what dark silicon means.

 

[Mauller]

That is not what dark silicon means.

Unused silicone then, but the whole point i was making still stands.

 

[bru]

Please enlighten me in your infinite wisdom as to how dedicated DP units are used when a GPU performs graphics processing? Because in the majority they aren't and remain inactive but still using power. Hence Dark Silicone as it is silicone that is not being used.

What ?????

all he did was snigger at "DP units are dark silicone"
he made no reference to what is or isn't used when graphics are processed.