Nvidia Unveils Pascal Gpu: 16gb Of Memory, 1tb/s Bandwidth

[sandboarder08]

Wonder what will be a bottleneck cpu?

[Jahramika]

No need for Nvlink other than to try and make a open platform closed to Nvidia. Software now does not even come close to using the full bandwidth of PCI 3.0 Bus. 

Edited November 17, 2015 by Godlygamer23

[Jerakl]

Both have drivers that take turns at being shitty so......

 

Yeah IK.

As far as I can tell right now it's Nvidia's turn to have worse drivers.

[PlayStation 2]

eww

[iBeast_M0de]

Actually Maxwell was an anomaly, I think only the Titan for a while then the 1080 will feature pascal and all others will be maxwell rebrand like 1070 = 980 and such.

It might by but I doubt it Nvidia has enough dev money to make a full new series Amd on the otherhand sadly doesnot have that kind of dev money

[don_svetlio]

Yeah IK.

As far as I can tell right now it's Nvidia's turn to have worse drivers.

Last 3 drivers are unusable by me - they are most definitely shitty

[Misanthrope]

It might by but I doubt it Nvidia has enough dev money to make a full new series Amd on the otherhand sadly doesnot have that kind of dev money

 

Yes they do have the money, but you gotta think that they haven't maxed out maxwell yet. Or in other words, yes in the top end but by the time they roll out pascal they should have a lot better yields on maxwell meaning that they should be able to save significantly on it and making the refresh better. AMD does this (and Nvidia too but to a lesser extend) not only to save money but because the performance advantage of newer gen products is also not linear: by the time you reach mid range cards you can save a bunch of money just by more aggressively clocking a previous gen chip like Hawaii to Grenada for example. Better binned chips and better clocks at lower TDP means they can save a lot more and have almost equal performance instead of using cut Fiji chips that are much more expensive for not that much more performance outside of top end cards. This should work the same for Nvidia

[-BirdiE-]

I NEED THIS!

[Jerakl]

Last 3 drivers are unusable by me - they are most definitely shitty

 

The one for SW:Battlefront nerfed my performance by about 30+%

[Glenwing]

It's amazing how many times the same information can just be republished again after a few months and people will get excited about the same thing... Again... All this information was released way back when the Titan X came out.

[don_svetlio]

It's amazing how many times the same information can just be republished again after a few months and people will get excited about the same thing... Again... All this information was released way back when the Titan X came out.

It's because "Nvidia!" - I guarantee if AMD were announcing something similar the response would be "OLD!"

[ACatWithThumbs]

Anyone has detail on pricing?

To early to tell.

We don't know currently how much the new tech costs and the price for the most part will depend on what AMD has on release.

You have to keep in mind that Pascal will destroy anything that's currently out by large margin so until AMD releases their 16nm GPU's Nvidia can pretty much price it how they want.

Just to put in perspective from what we know so far the flagship should look something like this:

Pascal:                                                Titan X:

16nm 5000-6000 Cuda Cores           28nm 3072 Cuda Cores

16GB 1/TB's                                        12GB 336GB/s

My guess is that the flagship will be around 2x as fast a 980Ti.

 

[VagabondWraith]

To early to tell.

We don't know currently how much the new tech costs and the price for the most part will depend on what AMD has on release.

You have to keep in mind that Pascal will destroy anything that's currently out by large margin so until AMD releases their 16nm GPU's Nvidia can pretty much price it how they want.

Just to put in perspective from what we know so far the flagship should look something like this:

Pascal:                                                Titan X:

16nm 5000-6000 Cuda Cores           28nm 3072 Cuda Cores

16GB 1/TB's                                        12GB 336GB/s

My guess will be that the flagship will be around 2x as fast a 980Ti.

 

 

If that's true, 4K might be my choice (over 1440p) after all. 4K is so (excuse my French) fucking beautiful that I really hope this will push that res.

[don_svetlio]

If that's true, 4K might be my choice (over 1440p) after all. 4K is so (excuse my French) fucking beautiful that I really hope this will push that res.

I really doubt DOUBLE the power of GM200 is going to be that easy. I'm skeptical.

[sandboarder08]

I really doubt DOUBLE the power of GM200 is going to be that easy. I'm skeptical.

Didn't they say 10X the performance of current gen. (Commercial)

http://cdn.wccftech.com/wp-content/uploads/2015/03/Screenshot-95.png

[VagabondWraith]

Didn't they say 10X the performance of current gen. (Commercial)

http://cdn.wccftech.com/wp-content/uploads/2015/03/Screenshot-95.png

Compute performance over Maxwell (since Maxwell is gimped FP64 wise). It doesn't mean gaming wise whatsoever.

[sandboarder08]

Compute performance over Maxwell (since Maxwell is gimped FP64 wise). It doesn't mean gaming wise whatsoever.

I'm guessing that most gains will be at the higher res. So at 1080 maybe 1.5x and 4k 2.5x. HBM is a game changer as amd saw.

[VagabondWraith]

I'm guessing that most gains will be at the higher res. So at 1080 maybe 1.5x and 4k 2.5x. HBM is a game changer as amd saw.

I'm def excited for next gen. I've already got $600 saved for 1440/4K monitor, Pascal flagship GPU and possibly a jump to the i7 6900K.

[ACatWithThumbs]

I really doubt DOUBLE the power of GM200 is going to be that easy. I'm skeptical.

It will be we might not see it release as the first card though if Nvidia wants to keep the larger cards for later.

Maxwell was faaaar slower than it was supposed to be because it was on 28nm instead of the planned 20nm.

We are now jumping two process shrinks at once the process size is almost cutting in half that alone will give us almost 2x as much space.

Percentage wise this is the largest jump from one process shrink to another in the last decade.

HBM 2.0 has over 3x the memory bandwidth of GDDR5 so the increase there is also obvious.

The only thing that will make or break it are the ROP's and Texture Units if those scale accordingly with the before mentioned specs we will double performance if not more.

If those do not scale accordingly we could see a result like in the FuryX where HBM is basically bottle necked by the ROP count.

And all of this doesn't even account for basic architecture optimizations as well as software optimizations.

 

[VagabondWraith]

It will be we might not see it release as the first card though if Nvidia wants to keep the larger cards for later.

Maxwell was faaaar slower than it was supposed to be because it was on 28nm instead of the planned 20nm.

We are now jumping two process shrinks at once the process size is almost cutting in half that alone will give us almost 2x as much space.

Percentage wise this is the largest jump from one process shrink to another in the last decade.

HBM 2.0 has over 3x the memory bandwidth of GDDR5 so the increase there is also obvious.

The only thing that will make or break it are the ROP's and Texture Units if those scale accordingly with the before mentioned specs we will double performance if not more.

If those do not scale accordingly we could see a result like in the FuryX where HBM is basically bottle necked by the ROP count.

And all of this doesn't even account for basic architecture optimizations as well as software optimizations.

 

nVidia is unlikely to gimp ROPs.... unlike AMD.

 

And I forgot about nVidia having to settle for 28nm instead of 20. They were banking on it being ready but it wasn't meant to be. Therefore, they had to restructure. People wondering where nVidia's "compute" performance went with Maxwell, well there ya go.

[AMD_INTEl wiztips]

Oh MA GAUD , Unreliable Information, NVLinks will never appear now in consumer age , Supercomputer uses it like the tatian Supercomputer . PCIE Gen 4 might be coming in soon .

[Glenwing]

I really doubt DOUBLE the power of GM200 is going to be that easy. I'm skeptical.

It does have twice as many transistors (8B to 17B), so double performance isn't that unreasonable.