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

[Boomstick777]

Thing is Nvidia are ahead in DX11 in 90% of things, and by the time DX12 launches Nvidia's new cards will likely be much better at DX12 and people will happily upgrade anyway. Whereas not so much the same
attitude towards the competition. Tis a shame for AMD but what can you do, Nvidia reads the market and it's marketing_brand power are just spot on. They will deliver best in class DX12 cards just in time for the games imho. AMD need to be on the money with the next lot of cards to stand any chance of regaining any serious market share.

 

[Mauller]

No idea what Polaris or Pascal will be having but interested to read what info you have about Polaris and ROVS and CR. I always like to keep updated.

It is more of a Given that ROV's will be supported. In fact Intel supported ROV's in hardware before Nvidia and were the ones to big it up in the first place. Order Independent Transparencies is the technical name for it which intel used for the feature.

Another funny thing is that the Dreamcast supported OIT in hardware.

conservative rasterisation will more than likely be part of the Primitive discard accelerator that was announced.

 

[Gregster]

It is more of a Given that ROV's will be supported. In fact Intel supported ROV's in hardware before Nvidia and were the ones to big it up in the first place. Order Independent Transparencies is the technical name for it which intel used for the feature.

Another funny thing is that the Dreamcast supported OIT in hardware.

conservative rasterisation will more than likely be part of the Primitive discard accelerator that was announced.

Well assumption aside, I hope you are correct and Nvidia have hardware ACE.

 

[CAT-THE-FIFTH]

Thing is Nvidia are ahead in DX11 in 90% of things, and by the time DX12 launches Nvidia's new cards will likely be much better at DX12 and people will happily upgrade anyway. Whereas not so much the same
attitude towards the competition. Tis a shame for AMD but what can you do, Nvidia reads the market and it's marketing_brand power are just spot on. They will deliver best in class DX12 cards just in time for the games imho. AMD need to be on the money with the next lot of cards to stand any chance of regaining any serious market share.

It could massively backfire on Nvidia. If the current AMD cards do a decent amount better through the whole range in DX12 titles(especially cards like the R9 390 and the R9 380),then you could end up with quite a few annoyed customers.

I give you an example yonks ago. Nvidia sold more FX series cards than ATI did for the 9000 series,but interestingly enough during the X800 days ATI reached parity with the 6000 series Nvidia cards in sales!! I suspect it might have been down to lots of annoyed FX users buying an ATI card,when their FX GPUs were not doing that well in DX9 titles(despite the Nvidia 6000 series being technically better than the X800 series as it supported DX9.0c).

 

[Orangey]

Did anyone make a Pasquale joke yet?

 

[Gregster]

Pascal will feature 4X the mixed precision performance, 2X the performance per watt, 2.7X memory capacity & 3X the bandwidth of Maxwell. Nvidia’s CEO went on to state that all in all Pascal is Maxwell times ten. All of this has just been revealed here at GTC. There’s a lot to digest here, so let’s break it down.

Read more: http://wccftech.com/nvidia-pascal-gpu-gtc-2015/#ixzz422g6X25o

 

[Rroff]

Hehe before anyone gets carried away the 10x is "CEO Maths" and basically he is saying that with all the enhancements in combination you have "ten times the GPU".

 

[Kaapstad]

Hehe before anyone gets carried away the 10x is "CEO Maths" and basically he is saying that with all the enhancements in combination you have "ten times the GPU".

+1

In practice I think the first mid range Pascal cards that turn up will be struggling to beat a GTX 980 Ti by a small amount.

 

[humbug]

Hehe before anyone gets carried away the 10x is "CEO Maths" and basically he is saying that with all the enhancements in combination you have "ten times the GPU".

yeah...

4X the mixed precision performance
2X the performance per watt
2.7X memory capacity
3X the bandwidth of Maxwell

He's being modest, its actually 11.7 times Maxwell

Looks impressive but again don't get too excited, "4X the mixed precision performance" is simply the by-product of HBM2, its 4x faster so the Floating Point performance will be 4x that of GDDR5. it has little to do with gaming performance, tho may help in Direct Compute 5 tasks.

 

[Mauller]

Hehe before anyone gets carried away the 10x is "CEO Maths" and basically he is saying that with all the enhancements in combination you have "ten times the GPU".

Technically pascals is 10x maxwell, in double precision compute. 1/32 for maxwell to 1/3 for pascal.

 

[Kaapstad]

Technically pascals is 10x maxwell, in double precision compute. 1/32 for maxwell to 1/3 for pascal.

For gaming this will actually hinder Pascal as it is not needed and will hold the card back.

 

[Rroff]

yeah...

4X the mixed precision performance
2X the performance per watt
2.7X memory capacity
3X the bandwidth of Maxwell

He's being modest, its actually 11.7 times Maxwell

Looks impressive but again don't get too excited, "4X the mixed precision performance" is simply the by-product of HBM2, its 4x faster so the Floating Point performance will be 4x that of GDDR5. it has little to do with gaming performance, tho may help in Direct Compute 5 tasks.

I highly suspect that tier for tier we are looking at approx +45% increase on average in performance in DX11 and older games and upto 2-3x the performance for games that use the latest DX11 extensions and/or DX12+.

 

[humbug]

Depends on how big they are going to make them ^^^ most of the performance increase will come from a node half the size, HBM2 will increase the performance somewhat but unless there are architectural improvements Pascal will be similar in performance per transistor to Maxwell. IMO don't expect 2x the FPS from a card half the size. maybe about <1.25 times from a card half the size

For gaming this will actually hinder Pascal as it is not needed and will hold the card back.

It will, i think its about time for separating workstation cards from gaming cards.

More power efficient and better overclocking without the compute engine getting in the way.

 

[Mauller]

For gaming this will actually hinder Pascal as it is not needed and will hold the card back.

They are adding the Double precision compute parts back to pascal again, so just like with kepler and AMD cards, even if it is not used it will still be in silicon.

The only reason Maxwell clocks so well is because it is stripped back to 32bit compute with a minimal amount of double precision. That removed a lot of silicon.

 

[Mauller]

Depends on how big they are going to make them ^^^ most of the performance increase will come from a node half the size, HBM2 will increase the performance somewhat but unless there are architectural improvements Pascal will be similar in performance per transistor to Maxwell.



It will, i think its about time for separating workstation cards from gaming cards.

More power efficient and better overclocking without the compute engine getting in the way.

It also depends on how the cards perform double precision, if it fuses two 32bit cores the it is more efficient then having dedicated double precision on the card. I think this is how AMD mostly does it considering how tahiti, hawaii and Fiji have 1/2 double precision. While Nvidia were more partial to having dedicated 64bit hardware. which gave them a lower overall double precision compared to the GCN parts.

ofcourse there still needs to be ancillary hardware for this fusing which adds extra transistors but it uses fewer overall compared to dedicated hardware.

 

[Kaapstad]

They are adding the Double precision compute parts back to pascal again, so just like with kepler and AMD cards, even if it is not used it will still be in silicon.

The only reason Maxwell clocks so well is because it is stripped back to 32bit compute with a minimal amount of double precision. That removed a lot of silicon.

+1

This is how they made the GTX 980 so competitive against the original Titans.

 

[Simon Belmont]

+1

In practice I think the first mid range Pascal cards that turn up will be struggling to beat a GTX 980 Ti by a small amount.

980Ti by 20%. Just my opinion.

 

[Mauller]

980Ti by 20%. Just my opinion.

With Finfett and the shrink, we might even see the first 2ghz clocked GPU. So it could be anywhere around 20% increase over a 980TI for mid pascal and mid polaris.

 

[D.P.]

They are adding the Double precision compute parts back to pascal again, so just like with kepler and AMD cards, even if it is not used it will still be in silicon.

The only reason Maxwell clocks so well is because it is stripped back to 32bit compute with a minimal amount of double precision. That removed a lot of silicon.

Fiji also has a dramatically reduced DP conure performance. AFAIk, Maxwell and Fiji actually have much close theoretical compute performance than Kepler and Hawaii. So AMD will also have to add back in aor of the removed compute silicon

 

[Mauller]

Fiji also has a dramatically reduced DP conure performance. AFAIk, Maxwell and Fiji actually have much close theoretical compute performance than Kepler and Hawaii. So AMD will also have to add back in aor of the removed compute silicon

In hardware, fiji is capable of 1/2 DP. but it is of course hobbled in software for consumer cards. Just check the Firepro versions of Tahiti, Hawaii. Both are 1/2 DP as well. Or at least the majority are if you check the highest end firepro cards.

The difference is that Maxwell does not have the DP silicon at all, its not just a matter of it being hobbled in software. It is the main reason why Kepler cards still fill the high end for Quadro and Tesla parts.

Edit-

Ooh, now that is dirty of Nvidia, i never noticed that they replaced all the Quadros with Maxwell models now. And they don't state any of the specs on their website for SP or DP compute performance. Only the Tesla cards state their Compute performances and as i mentioned they are all Kepler parts for the highest end models. They used to have the specs before for the Kepler Quadro cards, showing SP and DP compute.

None of the Professional Maxwell cards have their compute specs in easy to find places, i am still looking and cant find them. The Tesla K40 and K80 were directly on the products pages. Didn't even have to trawl through PDF's.