Nvidia Pascal Announcement

[fellix]

Looks like the clock-rate advantage of the new design and manufacturing process is the leading performance capital in Pascal. The power efficiency improvement comes naturally after that.
It's weird that there was not a single word about any advantages of double-speed FP16 support for graphics.

 

[Ext3h]

Actually 1070 is supposed to be faster than Titan X "in VR when using simultaneous multiprojection", not in general.

Yes, there were two slides. But the numbers Huan brabbled over and over again during the last 5 minutes were only "2x performance, 3x efficiency".

Maxwell can do this to 9 viewports, Pascal can do this to 16.

And now I'm trusting Huans last numbers even less. The presentation showed that they only needed/used 8 viewports for stereoscopic VR, (mostly) compensating the overhead for lens correction. So there is no reason why they couldn't just backport that approach to Maxwell for a fair comparison. In which case this should probably be the more plausible "1.3x" number for VR as well.

 

[Grall]

Jen-Hsun was lovably hammy as usual in his presentation. Pretty wicked hardware, I have to say. The 980 always felt like a gimped, cut-down Titan sold at a premium price to me, but the 1080 seems a much more performance-oriented chip in a way. The founder's edition - crazy priced as it is - in particular is pretty wicked.

Funny to see Nvidia calling out the PC platform's openness as one of its strengths, when NV has been a force for anything but openness at least ever since they introduced cuda, then physx and gameworks. I wonder if they're ever truly going to change their ways, I heard Jen-Hsun say something about open sourcing something at the end, but I was only half listening at that point p) so I'm not sure what exactly he was referring to - the surround sound APIs presumably...? Not all of gameworks, certainly?

I wouldn't mind a founder's 1080 or two, even with the blower cooler, as my small uATX casing isn't well suited for standard axial fan coolers. I can't do SLI with axial coolers - the cards just run too friggin hot.

It also seems NV has returned to a reference design with a backplate; if the unit shown on stage is representative of the final production model. Good!

 

[CSI PC]

So any ideas how they have implemented that Single Pass Stereo rendering?
Could be very interesting, and assuming seperate to SMT.
Cheers

 

[Ext3h]

So any ideas how they have implemented that Single Pass Stereo rendering?
Could be very interesting, and assuming seperate to SMT.
Cheers

https:// youtu.be/it3HVZMSBfY?t=1h3m14s

EDIT: Grrrr.... Stupid forum doen't let nme post a link to youtube with timestamp. Sorry, got to break the link to work around that.

 

[CSI PC]

On their website,
they show RotTR (DX12) being roughly 75% faster and Witcher 3 roughly 60-65% faster with the 1080 compared to 980.
Performance tab: http://www.geforce.co.uk/hardware/10series/geforce-gtx-1080
Which makes sense if it is faster than 980 in sli.
Cheers

 

[CSI PC]

https:// youtu.be/it3HVZMSBfY?t=1h3m14s

EDIT: Grrrr.... Stupid forum doen't let nme post a link to youtube with timestamp. Sorry, got to break the link to work around that.

Yeah and thanks, I did watch it but it sort of lightly explained how it is implemented within Pascal.
Curious what architecture-design improvements were required for this as it seems unique to Pascal and not with Maxwell.
Cheers

 

[Ext3h]

Yeah and thanks, I did watch it but it sort of lightly explained how it is implemented within Pascal.
Curious what architecture-design improvements were required for this as it seems unique to Pascal and not with Maxwell.
Cheers

I don't think it is. @MDolenc pointed out that the feature was already present in Maxwell. And 9 viewports as supported by Maxwell would have been enough for this specific implementation, the full 16 ones supported by Pascal are not even required.

So I suspect that Nvidia simply didn't want to show that even Maxwell did receive a performance boost from a backport, since that would have rendered that impressive "2x" statement invalid.

 

[Love_In_Rio]

If Fury X is 300 watts, anf AMD claims of 2xefficiency are true, wouldnt it be possible for a small Polaris 10 chip at 150-175 watts to be more powerful than a 1080 GTX?.

 

[Pressure]

That's a massive price hike!

From $549 to $699 ... we all know the only cards available for the foreseeable future will be the "Founders Edition".

 

[CSI PC]

I don't think it is. @MDolenc pointed out that the feature was already present in Maxwell. And 9 viewports as supported by Maxwell would have been enough for this specific implementation, the full 16 ones supported by Pascal are not even required.

So I suspect that Nvidia simply didn't want to show that even Maxwell did receive a performance boost from a backport, since that would have rendered that impressive "2x" statement invalid.

I have never seen it mentioned for VR relating to Maxwell, especially from developers who were critical of Maxwell *shrug*, this is different I assume to SMT.
None of us no for sure how it is implemented and I think some are assuming it is just driver-software, which it may be but prefer to know for sure as it would had been presented in the past (although I accept I may had missed that) rather than specifically with Pascal.
Anyone has information showing this being presented before with Maxwell or VRWorks+Maxwell?

Edit:
Ok it does seem linked to Pascal, this is from the VRWorks site:

Traditionally, VR applications have to draw geometry twice -- once for the left eye, and once for the right eye. Single Pass Stereo uses the new Simultaneous Multi-Projection architecture of NVIDIA Pascal-based GPUs to draw geometry only once, then simultaneously project both right-eye and left-eye views of the geometry. This allows developers to effectively double the geometric complexity of VR applications, increasing the richness and detail of their virtual world.

https://developer.nvidia.com/vrworks

Cheers

 

[Voxilla]

Close to nothing has been said about 1070's performance and I sincerely doubt it'll match a Titan X.

I'm not buying it neither be it only for the lack of bandwidth of 256 GB/s vs 336 GB/s
Wish Jen-Hsun Huang would come back down to earth and stop blowing things out of proportion like he's been doing for the last number of GPU introductions.

 

[trinibwoy]

At Maxwell launch they called it "multi-projection acceleration". I suppose that's technically different from simultaneous multi projection.

 

[Alessio1989]

sli bridges again.....

 

[sebbbi]

So any ideas how they have implemented that Single Pass Stereo rendering?
Could be very interesting, and assuming seperate to SMT.
Cheers

Oculus OpenGL extensions:
https://www.opengl.org/registry/specs/OVR/multiview.txt
https://www.opengl.org/registry/specs/OVR/multiview2.txt

Basically this extensions runs (part of) the VS and PS multiple times, once per viewport. A viewport index SV input has been added. You could have already implemented similar stuff yourself with geometry instancing (render 2 instances of everything) + VS output viewport array index (DX11.3 / 12 API and most DX 12 compatible GPUs). Output even instances to VP0 and odd to VP1. No extra draw calls (no extra CPU or GPU command processor cost).

Some (mobile) article about using the new Oculus extension:
http://malideveloper.arm.com/resources/sample-code/using-multiview-rendering/

 

[silent_guy]

From $549 to $699 ... we all know the only cards available for the foreseeable future will be the "Founders Edition".

Do we all know that? AFAIK, there was no such differentiation between model availability.

 

[Picao84]

IMO, another lackluster presentation by nVIDIA with extremely vague announcements and worse, no examples of performance with and without asynchronous compute, which probably means that nVIDIA indeed does not have a proper implementation of it. Performance numbers were thrown around in a very random fashion and most details about the chips were not presented at all. Then, the "Founders Edition" novelty silliness sounds like a PR way of hiding what may be low yields. It all just seemed sad and artificial, like something is seriously wrong and they know its a matter of time till it is discovered. At least this is my impression of it (and GP100 launch).

 

[CarstenS]

In the press release, they're touting "New asynchronous compute advances improve efficiency and gaming performance."
Rather curiously placed under the paragraph "Super Craftsmanship".

Yesterday, they weren't ready to give details about the difference between MPA and simultaneous multi projection apart from some vague flexibility thing that's not possible with maxwell, but is with pascal now. Doubt if there will be more nitty-gritty detail about this until a future date.

 

[CSI PC]

Oculus OpenGL extensions:
https://www.opengl.org/registry/specs/OVR/multiview.txt
https://www.opengl.org/registry/specs/OVR/multiview2.txt

Basically this extensions runs (part of) the VS and PS multiple times, once per viewport. A viewport index SV input has been added. You could have already implemented similar stuff yourself with geometry instancing (render 2 instances of everything) + VS output viewport array index (DX11.3 / 12 API and most DX 12 compatible GPUs). Output even instances to VP0 and odd to VP1. No extra draw calls (no extra CPU or GPU command processor cost).

Some (mobile) article about using the new Oculus extension:
http://malideveloper.arm.com/resources/sample-code/using-multiview-rendering/

Thanks Sebbi,
Does this mean though they expected it for future hardware rather than existing when discussing single pass rendering in context like the Pascal presentation:

On existing hardware, applications and drivers can realize the benefits of a single scene traversal, even if all GPU work is fully duplicated per-view.
But future support could enable simultaneous rendering via multi-GPU, tile-based architectures could sort geometry into tiles for multiple views in a single pass......

Thanks

 

[CSI PC]

At Maxwell launch they called it "multi-projection acceleration". I suppose that's technically different from simultaneous multi projection.

That was for creating multiple view ports though, not single pass rendering left and right displays.
Cheers