No DX12 Software is Suitable for Benchmarking *spawn*

Well, there's RTX 30 series being slower than RTX 20 series and GTX 1660's and the latter getting twice the FPS of RX 6000 series in traditional rendering which should raise an eyebrow or two.

 

All this tells me is that Turing put more emphasis on traditional geometry performance than Ampere does. Which makes sense given its an older architecture.

Clearly the the most important metric in this mesh shader test is the mesh shader performance.

 

The perf scaling between the SKUs definitely underlines how AMD and NV have been implementing geometry progressing. The RDNA cards are grouped in a narrow range, regardless of their salvage status, while NV hardware is going up and down predictably by the SM count.

 

There are several weird results there.

1. All RDNA2 GPUs are showing the same results basically with MS off and on hinting at the fact that they are limited by some common part in both cases - L3 cache maybe? I'd expect some gains here with future drivers, at least between 6800 and 6800XT+
   
2. Turing is faster than Ampere in "traditional" pipeline. It's an interesting result which may point to Ampere's being balanced differently in its geometry processing. Or maybe its drivers too.
3. All Turing chips with MS off are awfully close to each other - with the exception of 1650S which begs the question of why is it behaving so differently here? It gains +330% with MS on while 1660 gets only +75%.
4. Same can be said about Ampere with 3060Ti being on par with 3090 when MS are off. This doesn't look right from what we know of these chips specs. Then again maybe we are seeing the FF setup limitations here which MS should help to avoid?

So? It doesn't mean that these results are invalid. It begs the questions I've asked above though.

If these are actually how cards will stack up in games with MS as well then I wonder if 2060 level will be where games will stop with MS utilization as this seem to be inline with what consoles should be able to achieve. (Then again with PS5 not supporting MS at all it's a moot point anyway.)

 

It's kind of curious though, isn't it? At first I thought turing was better because it might clock higher and the raster units operate per clock. But 1660 besting ampere? This data aggregate is either not reliable or they've stripped back the fixed function raster units a bit, which would be odd considering it'll take a long time for the industry to transition away from the old vertex shader pipeline.

 

indeed. Mesh Shaders only leverage the CU/SMs for their processing.

 

But why is 6800=6900XT then?

 

The AMD results look weird. They're not scaling with the number of CUs, which you'd think they would. The whole point of the mesh shader pipeline is to take advantage of the width of the gpu and do things in parallel.

 

I don't know. Typically it's mesh shader controller -> Compute. I'm unsure if there is a bottleneck on the controller/command processor itself that the benchmark is pushing beyond what it's designed to handle and thus we see these results. At least, that's the best I can spit ball if these results are correct.

 

better answer here:

 

2080Ti is a great card people were selling it for as low as 450$, due to fear of value loss before Ampere launch.

 

Looking at the results for the 3080,the benefits are kind of absurd.

15.2 ms per frame to 1.66 ms per frame.

 

Yeah, I've suspected as much. RNDA2 seem to be really dependent on driver optimizations.

 

VRAM. 1650 is the only 4 GB card and I measured around 5 GB memory allocation on my system when the test was not using mesh shaders.

Once the test enabled Mesh shading, VRAM allocating dropped dramatically to around 1 GB =1650 is no more VRAM bottlenecked and gets a huge speedup from that and Mesh shaders which results into the 400% difference. Atleast that's my theory but it's pretty rock solid.

 

It could be occlusion culling is only implemented with mesh shaders on.
Do we know anything about how they do it?

 

Radeon Software 21.2.2 has been released already - Radeon™ Software Adrenalin 2020 Edition 21.2.2 Release Notes

 

I would guess they don't do any sophisticated occlusion culling with mesh shaders off (computer shader based etc). They're probably comparing Input assembler -> vertex shader -> rasterizer to amplification shader -> mesh shader -> rasterizer. So it's not the gain we'd expect to see in games that are already leveraging more advanced forms of frustrum and occlusion culling.

 

That looks more reasonable. I imagine it will scale with the number of CUs as well, instead of being the same result across the product line.

 

Notice that NVidia prefers the smallest possible workgroup size (32) and AMD prefers the largest possible (128/256 - 256 if on console). Makes a huge difference on AMD.