(GURU3D RUMOR) AMD Polaris 10 GPU To Offer Near 980 Ti Performance For 299 USD?

[Magee_MC]

HTFS uses HW conservative rasterization, which is unsupported by AMD, not tessellation.

Also there is nothing wrong with using tessellation to reasonable levels. But of course, tessellation is considered a gift from the devil or something on this forum.

If Polaris has fast(er) tessellation and conservative rasterization I imagine it will all be forgotten and/or forgiven

The billion dollar question that causes so much consternation is what is a reasonable level of tesselation.

 

[renderstate]

The billion dollar question that causes so much consternation is what is a reasonable level of tesselation.

In a traditional rendering pipeline triangles with less than ~10 pixels don't generally contribute much to the image and they are slow to shade anyway. With more exotic rendering pipelines, like Pixar REYES, shading only happens on vertices and triangles need to be smaller than a pixel for good image quality.

 

[ThatBuzzkiller]

Fast voxelization also requires conservative rasterization. Tessellation has nothing to do with it.

https://developer.nvidia.com/vxao-voxel-ambient-occlusion

VXAO doesn't need HW conservative rasterization. It can work on ANY DX11 GPU, a way to enhance it though would be to use Maxwell's fast geometry shader feature exposed in NVAPI or OpenGL as the spec shows below ...

https://www.opengl.org/registry/specs/NV/viewport_swizzle.txt

 

[ThatBuzzkiller]

Perhaps you should recheck it

While you can use tessellation in both cases the key feature that allows ray traced shadows (HTFS) and voxelization is conservative rasterization. Keep in mind that you can do CS in SW with a geometry shader, but AMD is not very good at geometry shaders. I bet Polaris will support CS in HW since it's part of the new DX12 features.

I can only imagine how happy people would be if AMD had introduced ray traced shadows in real-time in complex scenes. But once again, since in this case it's an NVIDIA innovation, well, it must be bad/evil.. worse, tessellated! Ahh.. Pavlov could have a written an essay based on certain automatic responses.

This is simply not true ...

In the case of a passthrough GS, Nvidia can be faster but in the case of vertex data amplification, AMD is known to be faster last time I checked ...

HFTS is not ray traced shadows, they are frustum traced ...

 

[Magee_MC]

In a traditional rendering pipeline triangles with less than ~10 pixels don't generally contribute much to the image and they are slow to shade anyway. With more exotic rendering pipelines, like Pixar REYES, shading only happens on vertices and triangles need to be smaller than a pixel for good image quality.

That makes sense. At a certain point you start losing your tesselation bang for the buck. From what I understand the tesselation objections from those with AMD cards is that GameWorks games have been tesselating at a sub pixel level.

 

[renderstate]

https://developer.nvidia.com/vxao-voxel-ambient-occlusion



VXAO doesn't need HW conservative rasterization. It can work on ANY DX11 GPU, a way to enhance it though would be to use Maxwell's fast geometry shader feature exposed in NVAPI or OpenGL as the spec shows below ...



https://www.opengl.org/registry/specs/NV/viewport_swizzle.txt

I previously mentioned conservative raster can be implemented in a GS, it's just slow, especially on AMD HW. You are right on conservative raster not being strictly necessary for voxelization, but only if you can live with cracks between primitives. The most accurate voxelization schemes require conservative raster (but I don't know if VXAO uses it or not). There is a discussion on the cracks/hole problem here: https://developer.nvidia.com/content/basics-gpu-voxelization

Since voxelization with or without cracks requires a GS, it won't run very fast on current AMD HW. NVIDIA HW instead can accelerate both the orthographic projection case required by voxelization and conservative raster case to avoid cracks.

 

[renderstate]

That makes sense. At a certain point you start losing your tesselation bang for the buck. From what I understand the tesselation objections from those with AMD cards is that GameWorks games have been tesselating at a sub pixel level.

When it's so dense it's a waste. On the other hand with the current tessellation pipeline is not always easy to locally control polygonal density. Adaptive tessellation schemes make the implementation more complex and come with their own set of problems, like geometry popping in/out of existence. If your GPU can digest lots of geometry without massively slowing down it can be easy to over-tessellate without even realizing it.

 

[renderstate]

This is simply not true ...



In the case of a passthrough GS, Nvidia can be faster but in the case of vertex data amplification, AMD is known to be faster last time I checked ...



HFTS is not ray traced shadows, they are frustum traced ...

I might be but of date with how AMD HW handles geometry shaders. IIRC they dump everything to memory and then they reload it. Is this still the case? Even if they handle the amplification case well the rest of the fronted might have some problems digesting lots of triangles. Again, this is something that has probably been addressed in Polaris.

Tracing frustums is actually more complex than tracing rays. In terms of image quality HTFS gives contact shadows virtually identical to ray traced hard shadows. I am not sure people understand what a huge deal this is. It's the beginning of the end for shadow maps.

 

[Magee_MC]

When it's so dense it's a waste. On the other hand with the current tessellation pipeline is not always easy to locally control polygonal density. Adaptive tessellation schemes make the implementation more complex and come with their own set of problems, like geometry popping in/out of existence. If your GPU can digest lots of geometry without massively slowing down it can be easy to over-tessellate without even realizing it.

I appreciate your response. I'll try to dig into the topic to see if I can understand it better. I think that we're more than a bit off topic here, so I'll just leave it alone for now.

Thanks!

 

[ThatBuzzkiller]

I previously mentioned conservative raster can be implemented in a GS, it's just slow, especially on AMD HW. You are right on conservative raster not being strictly necessary for voxelization, but only if you can live with cracks between primitives. The most accurate voxelization schemes require conservative raster (but I don't know if VXAO uses it or not). There is a discussion on the cracks/hole problem here: https://developer.nvidia.com/content/basics-gpu-voxelization

Since voxelization with or without cracks requires a GS, it won't run very fast on current AMD HW. NVIDIA HW instead can accelerate both the orthographic projection case required by voxelization and conservative raster case to avoid cracks.

As far as GS being slow on AMD it looks like I was clear in your subsequent post ...

In the voxelization pass you can lower reliance on the GS via multi-projection acceleration for Maxwell or accessing the vertex data on GCN uarch ...

 

[ThatBuzzkiller]

I might be but of date with how AMD HW handles geometry shaders. IIRC they dump everything to memory and then they reload it. Is this still the case? Even if they handle the amplification case well the rest of the fronted might have some problems digesting lots of triangles. Again, this is something that has probably been addressed in Polaris.

Tracing frustums is actually more complex than tracing rays. In terms of image quality HTFS gives contact shadows virtually identical to ray traced hard shadows. I am not sure people understand what a huge deal this is. It's the beginning of the end for shadow maps.

AMD does emit each generated vertex to the video memory but they do have extensions that allow them to do per-primitive work other than in the GS ...

The GS and the command processor have very little to do with each other. The only possible objection against AMD's command processor would be that it can't handle small draws very well compared to Nvidia and by small draws I don't mean small triangles ...

HFTS is great but it's still very different to ray traced shadows. Frustum tracing requires that you create your irregular z-buffer and then create 4 planes per triangle to do the point-in-frustum testing. Ray tracing is all about doing ray to triangle intersection tests. I also see FTS as a big deal since shadow maps have too many drawbacks like aliasing, acne, and other artifacts but that's not the only way to do it since some GPUs are powerful enough to straight up do ray traced hard shadows!

I like FTS and more games should have it since it's a step in the right direction but it is not totally perfect since aliasing can be exhibited alpha tested shadows ...

 

[Pottuvoi]

The billion dollar question that causes so much consternation is what is a reasonable level of tesselation.

The level you need and not be limited by it.
Currently the hardware tessellation has some pitfalls and 64 levels are not enough.
Also you have to pre-tesselate surface to use it or do your own compute shader to do most of the work.

 

[Denly]

Minority here, I am hoping they release a +/- $100 low power low profile P11 to give older SFF PC some new live. Just need it for multi monitors support that help on media consumption, speed up light photo/video editing.

 

[Phynaz]

What is AMD making on 14nm FF now?

Technically, nothing. :sneaky:

 

[wahdangun]

Technically, nothing. :sneaky:

Technically they already producing gpu, did u see the hitman demo with polaris gpu?

 

[kraatus77]

Technically, nothing. :sneaky:

Just like GP100

 

[airfathaaaaa]

Technically, nothing. :sneaky:

still trolling i see

 

[xpea]

Just like GP100

at least we already have pictures of gp104 in the wild but still really nothing from polaris

 

[caswow]

at least we already have pictures of gp104 in the wild but still really nothing from polaris

so actually we dont have anything from either of both D:

 

[airfathaaaaa]

at least we already have pictures of gp104 in the wild but still really nothing from polaris

and we have 2 live demos your point being?

 

[xpea]

and we have 2 live demos your point being?

I would rather say "2 highly controlled behind closed doors blind demos that nobody clearly know what happened" ()
And my point is simple: both Nvidia and AMD can't control the leaks when their GPUs are in the field (AIBs, manufacturing plants, etc). But up to know, we only saw pictures of Pascal stuff (heat sink design, dies), telling me that Nvidia is closer to launch than AMD. Nothing more, nothing less

 

[airfathaaaaa]

I would rather say "2 highly controlled behind closed doors blind demos that nobody clearly know what happened" ()
And my point is simple: both Nvidia and AMD can't control the leaks when their GPUs are in the field (AIBs, manufacturing plants, etc). But up to know, we only saw pictures of Pascal stuff (heat sink design, dies), telling me that Nvidia is closer to launch than AMD. Nothing more, nothing less

a picture of a pre production card is telling you that pascal is close? you are actually posting this 2 days before pascal paper launch? D:

we both know they gonna launch on june both companies there is no point of arguing about it

 

[xpea]

a picture of a pre production card is telling you that pascal is close? you are actually posting this 2 days before pascal paper launch? D:

we both know they gonna launch on june both companies there is no point of arguing about it

1/ Your hate against Nvidia deserves you, calm down ()
2/ How do you know Pascal will paper launch ? proof please.
and BTW, we still don't know what kind of event this Friday will be. Press event under NDA ? Technical announcement ? Official product launch ? I don't know, and neither you, so let's wait and see (peacefully)

 

[Phynaz]

still trolling i see

It's a joke.

What is your main account, so I know who doesn't have a sense of humor?

 

[Phynaz]

Technically they already producing gpu, did u see the hitman demo with polaris gpu?

AMD doesn't have factories, they don't produce anything but data. Get it?