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

[TwsT]

I am really looking forward to pascal. Started saving already

 

[Gregster]

I can see HBM being the reason it might not make Q1 2016 but hopefully HBM or I should say HBM2 will be in plentiful supply in a few months and that doesn't hold up production. Seeing the massive stutter delay of stock on Fiji has me slightly concerned.

 

[Mauller]

I can see HBM being the reason it might not make Q1 2016 but hopefully HBM or I should say HBM2 will be in plentiful supply in a few months and that doesn't hold up production. Seeing the massive stutter delay of stock on Fiji has me slightly concerned.

That delay is to do with the supplier and joiner of the interposers not having ramped up their production before the release of fiji. but they have ramped it up now so stock should start coming back in good numbers.

 

[TwsT]

and price should drop there is no point AMD dropping the price till supply is good. They may aswell milk it whist it's so choppy

 

[Ian Evey]

By the time fury x is easily available and prices drop hbm 3 will be out

 

[Telecaster]

By the time fury x is easily available and prices drop, time travel will be out

 

[muziqaz]

Orangey, 1080 non ti will eat 980ti alive. That is given. We have jump from 28nm to finfet 16nm plus hbm2 power savings. We might see another HD5870 jump from both manufacturers. If it was 28 to 20 nm then yeah I see 1080 being on par with titan x or 980ti or small jump.
The question is will nvidia create a miracle in putting all the new tech pieces together and making them work straight away

 

[Rroff]

Orangey, 1080 non ti will eat 980ti alive. That is given. We have jump from 28nm to finfet 16nm plus hbm2 power savings. We might see another HD5870 jump from both manufacturers. If it was 28 to 20 nm then yeah I see 1080 being on par with titan x or 980ti or small jump.
The question is will nvidia create a miracle in putting all the new tech pieces together and making them work straight away

If you took GM200 and did a direct shrink to 16nm FF+ (unfortunately don't think you can do a direct optical shrink) you'd get a ~45% performance increase straight up - assuming the issues with leakage are fixed and depending a little on final clock speeds and that is without doing anything to take advantage of any other advantages of the process. There are some unknowns though with Pascal in respect that there will be more compute considerations than Maxwell, there isn't much information on architecture changes that might have performance implications, yields and so on that could impact on clock speeds, etc.

 

[muziqaz]

If you took GM200 and did a direct shrink to 16nm FF+ (unfortunately don't think you can do a direct optical shrink) you'd get a ~45% performance increase straight up - assuming the issues with leakage are fixed and depending a little on final clock speeds and that is without doing anything to take advantage of any other advantages of the process. There are some unknowns though with Pascal in respect that there will be more compute considerations than Maxwell, there isn't much information on architecture changes that might have performance implications, yields and so on that could impact on clock speeds, etc.

That's what I am saying, we basically are skipping one process node. There is a lot of more space for transistors.
It is basically like coming from SFF case with single card to full tower case with WS size motherboard and 4-5 pci-e slots

 

[Rroff]

Would be interesting too see what the difference would be taking GM200 and porting it to HBM as well as the shrink and what kind of potential that would have for performance scaling on its own clock speeds/extra SPs, etc.

 

[Orangey]

We're not skipping a node, the interconnect wiring is still 20nm. 16nm is limited by that.

http://semiengineering.com/will-7nm-and-5nm-really-happen-2/

 

[D.P.]

Orangey, 1080 non ti will eat 980ti alive. That is given. We have jump from 28nm to finfet 16nm plus hbm2 power savings. We might see another HD5870 jump from both manufacturers. If it was 28 to 20 nm then yeah I see 1080 being on par with titan x or 980ti or small jump.
The question is will nvidia create a miracle in putting all the new tech pieces together and making them work straight away

Ignore Orangey, he is a rapid fanboy, just look at his post. It starts out with compelte rubbish about the 780ti to 980 jump being insignificiant. He is the kind of person who will then say that Nvidia purposely gimped 780 drivers as a emans of explaining the eprformance difference. For some reaons it is veyr hard for AMD fanboys to appreciate the performance and efficiency gains Nvidia achieved with Maxwell, while at the same tiem they will praise Fiji depsite the fact it is significantly less efficient than Maxwell despite usign HBM.

 

[muziqaz]

We're not skipping a node, the interconnect wiring is still 20nm. 16nm is limited by that.

http://semiengineering.com/will-7nm-and-5nm-really-happen-2/

yes we are everyone cuts corners all over the place, and names their processes whatever they like, but still the jumps will be much much bigger than we used to have before going from node to node, or if we went to 20nm instead

 

[muziqaz]

Ignore Orangey, he is a rapid fanboy, just look at his post. It starts out with compelte rubbish about the 780ti to 980 jump being insignificiant. He is the kind of person who will then say that Nvidia purposely gimped 780 drivers as a emans of explaining the eprformance difference. For some reaons it is veyr hard for AMD fanboys to appreciate the performance and efficiency gains Nvidia achieved with Maxwell, while at the same tiem they will praise Fiji depsite the fact it is significantly less efficient than Maxwell despite usign HBM.

please don't fall over yourself while typing these comments remember to breathe as well, thanks

 

[Mr Oblong]

I couldn't wait for Pascal (was on a 6950!), so I hopped to a used MSI 290 Gaming (£130, pretty good I thought) to tide me over.

It's only a stopgap though, too many things are lining up to make (big, little, whatever) Pascal awesome.

 

[Jono8]

We're not skipping a node, the interconnect wiring is still 20nm. 16nm is limited by that.

http://semiengineering.com/will-7nm-and-5nm-really-happen-2/

Even so, the jump we got from a 780Ti to a 980Ti was on the SAME node without any new memory tech such as HBM.

Pascal will be fast.

 

[Orangey]

It says a 2x density increase. You would expect more than that for a 2 node jump.

I think you're all setting yourselves up for a big disappointment.

I don't think ANY GPU released until say 2017 Q2 is going to be particularly impressive for high-end owners.

 

[Mauller]

The last i heard, 16/14nm Finfet are just marketing names for 20nm using FinFets.

So it is only a single node difference, but the change to finfets is like an extra node change per se. Regardless what node the finfets are used on.

 

[Mr Oblong]

It says a 2x density increase. You would expect more than that for a 2 node jump.

I think you're all setting yourselves up for a big disappointment.

I don't think ANY GPU released until say 2017 Q2 is going to be particularly impressive for high-end owners.

Give in to the hype, embrace the (currently hypothetical) power!!!

 

[Orangey]

What have finfets given us really? 5 years of stagnation.

Their benefits can only be felt by phone chips.