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

[DarrenM343]

This ties in with the rumours that consumer Pascal is a die shrunk Maxwell. A rebrand if you will

Note: the above statement may or may not be based on actual facts.

Die shrink not exactly a rebrand as that would be like for like but lets hope it's more than that .

 

[Harlequin]

GP106 with 128bit ram??

 

[Mauller]

Plus AMD has a marketing department - that is a new one for me!

With their current dGPU share of sales, they need to make all the buzz and hype they can.

 

[AllBodies]

So around 300mm2 on 16NM - that could be around GTX980TI level performance at least(or a bit quicker) since I expect the card will have higher stock clockspeeds.

Hopefully, though they've only managed 1.88x the transistors per mm2 instead of 2x.

So at 300.5mm2 (half 980 Ti exactly), they'd need a 6.4% IPC improvement to match the 980 Ti.

They could get around this by just making it 350-400mm2 though. Then it'd best the 980 Ti easily.

 

[Orangey]

Hopefully, though they've only managed 1.88x the transistors per mm2 instead of 2x.

So at 300.5mm2 (half 980 Ti exactly), they'd need a 6.4% IPC improvement to match the 980 Ti.

They could get around this by just making it 350-400mm2 though. Then it'd best the 980 Ti easily.

Yeah so easy to just throw another 100mm2 at it in first year of production.

 

[Jono8]

Hopefully, though they've only managed 1.88x the transistors per mm2 instead of 2x.

So at 300.5mm2 (half 980 Ti exactly), they'd need a 6.4% IPC improvement to match the 980 Ti.

They could get around this by just making it 350-400mm2 though. Then it'd best the 980 Ti easily.

Remember the 980 has 2048 cores and the 780ti had 2880 cores and the 980 is still 15%ish quicker.

The fact it is 300mm means nothing. It could still decently beat a 980ti if we get a similar per core performance improvement with pascal.

 

[SimBy]

Remember the 980 has 2048 cores and the 780ti had 2880 cores and the 980 is still 15%ish quicker.

The fact it is 300mm means nothing. It could still decently beat a 980ti if we get a similar per core performance improvement with pascal.

980 has much higher core clock.

 

[humbug]

980 has much higher core clock.

970 only has 1664 shaders, that is 70% more shaders to the 780TI.

@ 1178Mhz the 970 is easily a match for the 780TI @ 928Mhz.

I doubt a clock speed increase of 25% makes up for the missing shaders, Maxwell is just architecturally different to Kepler. The 970 looks almost like a GTX 680/770 and yet it blows those out of the water by a huge margin.

 

[Rroff]

970 only has 1664 shaders, that is 70% more shaders to the 780TI.

@ 1178Mhz the 970 is easily a match for the 780TI @ 928Mhz.

I doubt a clock speed increase of 25% makes up for the missing shaders, Maxwell is just architecturally different to Kepler.

There are some cases where no amount of overclocking Kepler can make up for the advantages in the Maxwell architecture - advances to stuff like dynamic parallelism and instruction scheduling can make Maxwell upto 35% faster in an otherwise like for like situation though realistically that benefit is more like 10-15% in situations where it manifests in games.

 

[humbug]

There are some cases where no amount of overclocking Kepler can make up for the advantages in the Maxwell architecture - advances to stuff like dynamic parallelism and instruction scheduling can make Maxwell upto 35% faster in an otherwise like for like situation though realistically that benefit is more like 10-15% in situations where it manifests in games.

Yeah, Maxwell has much better Delta Colour Compression, Tessellation throughput, that Instruction Scheduling also allows it to handle Draw Calls more efficiently so its less prone to CPU bottlenecking, it also has a more powerful Shading Engine so nothing is equal to Kepler.

To use Intel's evolution model its very much a 'Tock' architecture, a very good one at that. far more 'Tock' than anything Intel have ever achieved.

 

[bfar]

Does GP104 dump some of the compute functionality found in GP100? If so, it might make up some ground on that alone.

 

[DmitryD]

Does GP104 dump some of the compute functionality found in GP100? If so, it might make up some ground on that alone.

Most certainly there are very few FP64 cores.

 

[TheRealDeal]

Does GP104 dump some of the compute functionality found in GP100? If so, it might make up some ground on that alone.

I doubt it's to much of a factor. The gtx980 stacks up pretty well against the Ti. It's pretty much where it should be when you look at the specs. I don't know enough to rule it out but I don't think it's the issue while having a look at the numbers.

 

[Mauller]

They will more than likely dump all of or the majority of the DP shaders per cuda block. Most likely having 1/32 like with maxwell since it is a waste of die space for consumer parts, 1/3 of a wafer would be wasted money if they bept 1/2 DP in consumer parts.

 

[AlamoX]

i think ppl expect too much of the first line up, it took amd and nvidia 5 years of refined 28nm process to get to what we have now, if you wanna compare the up coming line up to existing one, pick the first one released on 28nm, not maxwell

 

[Jono8]

i think ppl expect too much of the first line up, it took amd and nvidia 5 years of refined 28nm process to get to what we have now, if you wanna compare the up coming line up to existing one, pick the first one released on 28nm, not maxwell

Ok. The 680 which was medium kepler and nvidia's first 28nm gpu was loads faster than the 580 which was big Fermi.

 

[Harlequin]

Ok. The 680 which was medium kepler and nvidia's first 28nm gpu was loads faster than the 580 which was big Fermi.

hmmm and when someone on here asked that very same question:

https://forums.overclockers.co.uk/showthread.php?t=18391785

the answer was `no not worth it`.....

not the sort of answer for something `loads faster`

 

[lumik]

hmmm and when someone on here asked that very same question:

https://forums.overclockers.co.uk/showthread.php?t=18391785

the answer was `no not worth it`.....

not the sort of answer for something `loads faster`

Funny thing.
GTx980TI is only 23% faster than GTX980(stock vs stock) yet is here treated like some sort of miracle card.

GTX680 was faster than GTX580(stock vs stock) by 30% and people dont recomanded upgrade.

 

[andybird123]

People don't recommend based on stock for stock, the forum is called "overclockers uk" for a reason

 

[Jono8]

hmmm and when someone on here asked that very same question:

https://forums.overclockers.co.uk/showthread.php?t=18391785

the answer was `no not worth it`.....

not the sort of answer for something `loads faster`

Well it was about 30% quicker so thats a pretty decent jump. My point was in response to the guy who said to compare the first 28nm card to the first 16nm card which showed you could still easily get a significant boost from a full fat card to a cut down smaller die card but on a new process.

As the other guy said above me, people recommend the 980ti left right and centre for people on 290xs/980s etc and that is only about 30% faster than the tier below it.