Nvidia Pascal Announcement

[Ryan Smith]

Seems like ppl believe same day as FE:
https://forum.beyond3d.com/posts/1915242/

Which is weird, because one of the reasone cited for the price premium was TTM.

If the mounting holes are the same as GTX 980 (Ti), then partners can just take their existing coolers designed for those reference PCBs and strap them on to the 1080, so the TTM for semi-custom cards is very low. None the less, I'm taking a wait and see approach, as this still takes longer than getting the FE cards out the door.

Edit: Yeah, all of the holes on the outer rim of the card are identical, as is GPU and DRAM placement. So as long as your heatsink gives appropriate clearance over the entire right side of the card, then you can just reuse your cooler.

https://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_980/images/front.jpg
https://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_1080/images/front.jpg

 

[CarstenS]

While I obviously (have to) agree there, it's kinda strange anyhow, since Nvidia must've known that as well and one of the scarce reasons for a premium priced FE just evaporates.

 

[xEx]

Not just one...THE reason for the premium price just disappear.

 

[Razor1]

Unless ya really want that hunk of aluminum , would be good to bludgeon one's self with.

 

[Malo]

Well [H] covered one aspect which has some validity to it

Founders Edition for the life of the GPU

Some of the big push for this product line seems to have come from Kelt Reeves, owner of Falcon Northwest. He was not shy at all in telling me that he had actively lobbied for this Founders Card product line. His position is this: he wants a video card product that does not change at all in terms of quality or specfication over the life of the GPU. Also Kelt explained that closed face coolers that only exhaust at the I/O shield end of the card are a must for all of his specialty SFF system builds. If you look at it from the perspective of FalconNW or the likes of Maingear, having a product that they can qualify once and sell for years is very important. Also, going back to the craftsmanship aspect, Kelt explained that he very much wanted Founders Edition quality builds in every one of his systems. It simply takes away a lot of concern about quality. Also, I would assume from our conversation that he will be purchasing these cards directly from NVIDIA and not an AIB partner.

 

[Razor1]

that actually makes sense!

 

[CarstenS]

True, but that's pure OEM/SB value, in my book at least no reason for heavily increased end user pricing.

 

[Razor1]

yep

 

[xEx]

Well there is a reason....selling a 320mm chip for 700 dollars is a very good one.

 

[MDolenc]

You are right, it's not much different. Except that back then VRWorks only did multi-resolution, means different viewport, but only a single projection matrix. Going by the GL_AMD_vertex_shader_viewport_index extension, AMD features something just as capable as Nvidias fast geometry shader path, so there shouldn't be any actual difference in performance.

But you don't need to allow arbitrary quads for viewports in order to do lens matched shading. You can just subdivide into 12 or 16 rectangles right away, and just apply an adjusted projection matrix to the vertex dispatched by the GS to each single viewport. Apart from trading masking for rectangular viewports, that's exactly the same as what Nvidia now showed.

(Well, you can't divide into more than 9 viewports on Maxwell, because that's the limit for the viewport array size on Maxwell. Interesting enough, GCN allowed 16 viewports from the beginning, just as Pascal does now.)

I'll have to revise my earlier post on this, since it contained quite a few errors. Though in my defense we are talking about something that's part of NV extension mechanics...

The point is that NV allows viewport bitmask, not just viewport index. Yes AMD has GL_AMD_vertex_shader_viewport_index and can send triangle to specified viewport without invoking geometry shader. But the problem is that it can't multicast triangle to multiple viewports (without invoking GS). In case of multi resolution rendering that NV demoed on Maxwell some triangles will most definitely span viewports. This means they will have to be sent to multiple viewports and Maxwell can do that. And it can do that for all 16 API defined viewports (and to all them at the same time) not just 9 as I claimed in my previous post.
It can also swizzle coordinates which allows it to render cube map in a single pass in this fast GS regime. It also allows you to discard triangle my setting viewport mask to 0.

With regards to Pascal there is actually not any actual information on how you will go around implementing this single pass stereo and other stuff they demoed. But it's easy to imagine that instead of 16 swizzle modes for viewports they extended this to 16 matrices (or that they are read from some constant buffer slot, which is actually more likely).

 

[3dilettante]

Isn't this within margin of error? The difference is less than 3%..

To follow up on this, it was clarified that async is still disabled, which can point to inter-run variation or some other non-compute factor nudging the needle.

 

[xEx]

Tear-Down of a 1080p. It has 1 million screws...g-zus.

 

[CSI PC]

Well [H] covered one aspect which has some validity to it

Indeed, but for the price increase they could had even improved on the design.
Take a look at the ASUS GeForce GTX 980 Matrix Platinum and how well it encloses the heat to ensure it all goes out through the exhaust, unfortunately the 1080 reference design has some leakage at the top of the card.

Reference 1080:

Asus 980 Matrix Platinum:

Links to both reviews:
http://www.guru3d.com/articles_pages/nvidia_geforce_gtx_1080_review,10.html
http://www.guru3d.com/articles_pages/asus_geforce_gtx_980_matrix_platinum_review,9.html

Cheers

 

[Ext3h]

But it's easy to imagine that instead of 16 swizzle modes for viewports they extended this to 16 matrices (or that they are read from some constant buffer slot, which is actually more likely).

If they did just lens matched shading for one eye, a viewport array with bitmask indexing would had been sufficient, assuming that they've extended both the "swizzle" as well as the viewport transformation matrix past clipping and scissoring to arbitrary matrices. But they didn't stop there. They used a different projection matrix for each eye as well.

We'll probably have to wait till the used DX11 extension is properly documented before we know for sure what Nvidia did. But so far it still smells as if they used the GS stage for vertex duplication. (Respectively I suspect that even using a bitmask to index the viewports internally maps to an appendix to the GS, unrolling the bitmask into individual vertices with full numerical indizes, even though I would need to benchmark that to confirm that...)

 

[renderstate]

You are right, it's not much different. Except that back then VRWorks only did multi-resolution, means different viewport, but only a single projection matrix. Going by the GL_AMD_vertex_shader_viewport_index extension, AMD features something just as capable as Nvidias fast geometry shader path, so there shouldn't be any actual difference in performance.

Unfortunately for AMD this is not the case since their extension only outputs to one viewport, while NV_viewport_array2 allows to output a triangle to many viewports at once.

But you don't need to allow arbitrary quads for viewports in order to do lens matched shading. You can just subdivide into 12 or 16 rectangles right away, and just apply an adjusted projection matrix to the vertex dispatched by the GS to each single viewport. Apart from trading masking for rectangular viewports, that's exactly the same as what Nvidia now showed.

Projection matrices are not associated to viewports. If you do change projection on a per viewport basis then you can't use the fast GS path on Maxwell. I suspect LSM, which I wouldn't characterizer as a non-rectangular viewport (I'm pretty sure the viewport is still a rectangle ), allows you to still use the fast GS path and a single projection matrix, or perhaps two, with single pass stereo (not sure of how that works yet..)

 

[Voxilla]

Google designing custom ASICs for deep learning.
http://www.eetimes.com/document.asp?doc_id=1329722&

Claiming 10 times improvements for performance per Watt
https://cloudplatform.googleblog.co...-machine-learning-tasks-with-custom-chip.html

Not much detail.

 

[3dilettante]

The eetimes article seems to rule out Pascal being involved.

 

[orcinus]

If the mounting holes are the same as GTX 980 (Ti), then partners can just take their existing coolers designed for those reference PCBs and strap them on to the 1080, so the TTM for semi-custom cards is very low. None the less, I'm taking a wait and see approach, as this still takes longer than getting the FE cards out the door.

Edit: Yeah, all of the holes on the outer rim of the card are identical, as is GPU and DRAM placement. So as long as your heatsink gives appropriate clearance over the entire right side of the card, then you can just reuse your cooler.

https://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_980/images/front.jpg
https://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_1080/images/front.jpg

I'm looking at those pictures and ... am I blind? Holes differ in the 2nd column from the right, where a middle hole is completely off to the right on 1080's PCB.

To follow up on this, it was clarified that async is still disabled, which can point to inter-run variation or some other non-compute factor nudging the needle.

Hmm, when they plan to enable it? ATM the result is very bad.

 

[Ryan Smith]

I'm looking at those pictures and ... am I blind? Holes differ in the 2nd column from the right, where a middle hole is completely off to the right on 1080's PCB.

It's the top and the bottom that matter. 3rd party open air coolers typically don't use the mounting hole on the right, since they're one flat heatsink extending to the end of the card (and often beyond). It's really only for blowers, which need it to secure the far end of the shroud.

 

[Jawed]

Google designing custom ASICs for deep learning.
http://www.eetimes.com/document.asp?doc_id=1329722&

Claiming 10 times improvements for performance per Watt
https://cloudplatform.googleblog.co...-machine-learning-tasks-with-custom-chip.html

Not much detail.

When Google's monthly revenue is more than NVidia's annual revenue, ASICs become possible and fucking about with GPGPU becomes history.