AMD Radeon RDNA2 Navi (RX 6500, 6600, 6700, 6800, 6900 XT)

[Scott_Arm]

Right. Higher resolutions put more pressure on bandwidth and fillrate. So if AMD has Ampere beat on the geometry frontend and fillrate they have a chance to sweep the board. I don’t know of any really ALU bound games.

Ray tracing?

 

[digitalwanderer]

Lets hope amd dont shoot themselves in the foot by releasing godawful drivers

From your mouth to Terry Makedon's ears!

 

[troyan]

Right. Higher resolutions put more pressure on bandwidth and fillrate. So if AMD has Ampere beat on the geometry frontend and fillrate they have a chance to sweep the board. I don’t know of any really ALU bound games.

Crysis Remastered with software raytracing. A 3090 is 3.8x faster than a 5700XT: https://www.purepc.pl/test-wydajnosci-crysis-remastered-czy-mi-pojdzie-zalezy-co-masz?page=0,14

 

[Scott_Arm]

I would expect UE5 is very likely going to be bottlenecked by ALU.

 

[iroboto]

I would expect UE5 is very likely going to be bottlenecked by ALU.

lumen sounds exceptionally heavy on ALU. And Epic did mention it was Lumen that was the greatest factor for performance and not their Nanite microgeometry

 

[Digidi]

I would expect UE5 is very likely going to be bottlenecked by ALU.

Fun Fact UE5 doing the most rasterisation on shader because bottleneck of frontend

 

[Scott_Arm]

lumen sounds exceptionally heavy on ALU. And Epic did mention it was Lumen that was the greatest factor for performance and not their Nanite microgeometry

I'm curious about mesh shader benchmarks. Besides nvidia asteroids, I don't know of any, but I'm not sure if that's directx or if it's vulkan with Nvidia proprietary extensions. I'd really like to get a sense of how mesh shaders will change front-end performance since you can throw compute at coarse and fine-grained culling and tessellation.

 

[Putas]

From your mouth to Terry Makedon's ears!

Oh god, is he still there? I cannot believe he survived 2004.

 

[trinibwoy]

Ray tracing?

I don’t see why hardware assisted raytracing would be ALU heavy in Nvidia’s case.

Crysis Remastered with software raytracing. A 3090 is 3.8x faster than a 5700XT: https://www.purepc.pl/test-wydajnosci-crysis-remastered-czy-mi-pojdzie-zalezy-co-masz?page=0,14

That’s an interesting data point. Are you sure it’s all software RT? I thought it’s also using hardware RT on Nvidia via some Vulkan + DirectX interface.

 

[tsa1]

It'd be interesting to compare Sienna's performance in Vantage fillrate benchmark (2nd feature test) with the previous generations. Something got wrong after Hawaii, it started to scale only with VRAM clocks (unlike contemporary nV gpus which were always GPU-clock bound)

 

[iroboto]

I'm curious about mesh shader benchmarks. Besides nvidia asteroids, I don't know of any, but I'm not sure if that's directx or if it's vulkan with Nvidia proprietary extensions. I'd really like to get a sense of how mesh shaders will change front-end performance since you can throw compute at coarse and fine-grained culling and tessellation.

Considering how different a mesh shader front end would be; would any game really want to support both?
I think once you go Mesh Shaders/Primitive Shader pipeline, I don't think you'll support older pipelines.
Its the same if you go gpu based submission, you're probably not going to want to support non-gpu based submission anymore.

 

[Erinyes]

If the 3080 cannot utilize it's bandwidth then neither can Turing and Pascal cards, since their bandwidth deltas correspond with perfomance deltas.
Only the 3090 has somewhat excessive bandwidth.

The context of the discussion was the 3080 vs the 3070, where the 70% additional bandwidth only translates into 26-32% more performance. The extra bandwidth does have a power cost and is certainly one of the reasons for the lower power efficiency of the 3080.

 

[Scott_Arm]

Considering how different a mesh shader front end would be; would any game really want to support both?
I think once you go Mesh Shaders/Primitive Shader pipeline, I don't think you'll support older pipelines.
Its the same if you go gpu based submission, you're probably not going to want to support non-gpu based submission anymore.

I think you can convert meshlets to meshes and then have both pipelines. Existing engines already have the vertex shader pipline in place. Mesh shaders are really new so not sure how practical that is.

 

[iroboto]

I think you can convert meshlets to meshes and then have both pipelines. Existing engines already have the vertex shader pipline in place. Mesh shaders are really new so not sure how practical that is.

Good point.
I just recall a point that stuck with me, and that Sebbbi knew async compute culling etc, could all outperform what we have, but no one wanted to do it because of lack of support and he was uninterested in supporting 2 separate pipelines (not the same as supporting 2 separate platforms)

I would assume they would just drop support for 1 because the mesh/primitive shaders are not only faster, but are functionality more versatile in terms of working with geometry.

 

[Scott_Arm]

I don’t see why hardware assisted raytracing would be ALU heavy in Nvidia’s case.



That’s an interesting data point. Are you sure it’s all software RT? I thought it’s also using hardware RT on Nvidia via some Vulkan + DirectX interface.

I don't think the Cryengine RT solution uses RT hardware. I think it's purely software. Otherwise 5700XT wouldn't be beating a 2060 here.

https://www.tweaktown.com/articles/...hmark-tool-tested/index.html#Benchmarks-1440p

 

[chris1515]

I don't think the Cryengine RT solution uses RT hardware. I think it's purely software. Otherwise 5700XT wouldn't be beating a 2060 here.

https://www.tweaktown.com/articles/...hmark-tool-tested/index.html#Benchmarks-1440p

The last version use ta mix of SVOGI and triangle based raytracing

https://www.cryengine.com/news/view...-ray-traced-reflections-in-cryengine-and-more

 

[Scott_Arm]

Good point.
I just recall a point that stuck with me, and that Sebbbi knew async compute culling etc, could all outperform what we have, but no one wanted to do it because of lack of support and he was uninterested in supporting 2 separate pipelines (not the same as supporting 2 separate platforms)

I would assume they would just drop support for 1 because the mesh/primitive shaders are not only faster, but are functionality more versatile in terms of working with geometry.

I think that's the main issue with mesh shaders. You have to support gpus like RDNA1, GCN, Vega and Pascal for a while longer. There's really no choice but to have two pipelines until you can drop those. Or you do what UE5 is doing and just roll your own software rasterizer that can work across any dx12/vulkan capable gpu.

 

[Scott_Arm]

The last version use ta mix of SVOGI and triangle based raytracing

https://www.cryengine.com/news/view...-ray-traced-reflections-in-cryengine-and-more

Yah, this link is a lot more clear. They do use RTX/DXR now and fall back to voxels on non-capable hardware.

 

[chris1515]

Yah, this link is a lot more clear. They do use RTX/DXR now and fall back to voxels on non-capable hardware.

No they mix Voxel and mesh based tracing the fallback is not SSR but Voxel cone tracing when mesh based tracing is not needed. For very little object like bullet case they used mesh based raytracing all the time. I think many engine will do like this in the future and not using SSR as a fallback. And you can use voxel for global illumination too.

One of the important components of any ray tracing implementation is the ray intersection acceleration structure. Our SVOGI (Total Illumination) system already contained what was necessary, so it was a relatively straightforward step to add the data required for ray tracing. In the current implementation, for every voxel, we store a reference to overlapping triangles, plus the usual information like albedo, opacity, and normal data.

Merging voxel and ray tracing data provides great flexibility. For example, for diffuse rays, true mesh tracing is needed only near the beginning of the ray but for the rest of the ray we can use more efficient cone tracing without any visual artifacts. A similar optimization works with specular rays. Only smooth and clean surfaces like mirrors require true mesh ray tracing. Most low-gloss, less shiny surfaces can be traced much faster simply by tracing voxels, which will achieve the same or better visual output. Another important part of most real-time ray tracing implementations is de-noising, which already exists in CRYENGINE.

Part of the process was delivering proper support for ray hit shading. Based on our existing deferred shading solution we implemented an additional render pass for the shading of rays, with the support of every possible light source we have. By doing this, we make sure that we get proper shading in reflections. We also added support for particle system ray tracing, allowing the reflection of particle effects.

 

[Putas]

The context of the discussion was the 3080 vs the 3070, where the 70% additional bandwidth only translates into 26-32% more performance. The extra bandwidth does have a power cost and is certainly one of the reasons for the lower power efficiency of the 3080.

I completely agree. A 3080-like card could have been easily more efficient. Just wanted to point out how broad is the implication of the conclusion of it having excessive memory bandwidth.