No DX12 Software is Suitable for Benchmarking *spawn*

Since 3dMark was mentioned my mboard bios has this option
3DMark01 Enhancement Allows you to determine whether to enhance some legacy benchmark performance. (Default: Disabled)

Anyone have any idea what it does?
edit: according to a youtube video I found it lowers the score by 41. The p.c scored about 8500 at 4k with it disabled
 
Last edited:
I’d guess if you’re not VRAM limited you won’t see much improvement except maybe from some bandwidth saving if any of the test is bandwidth limited.
This test has nothing to do with VRAM, as it's using the Texture Space Shading part of Sampler Feedback, not the more interesting streaming technique!

Performance varies greatly though depending how the base performance of the benchmark is. I got nearly a 20% more performance while my 2060 laptop was on battery.
 
What GPU?

3090.

I'd imagine that the more shading power a GPU has the lower the gains from using texture space shading will be.

It's not a benchmark of texture space shading vs regular shading. It's a benchmark of how much faster texture space shading is when using hardware sampler feedback. But yeah the whole point of texture space shading is to reduce the amount of shading in each frame so if thatisn't a bottleneck you won't see a significant benefit. Also if their "software solution" is bandwidth heavy that could also account for some of the difference.

"In the first pass, we shade using texture space shading without sampler feedback, using a software solution developed in-house to determine the sampled texels along with a matching software sampler in the screen space sampling pass. The second passes uses the sampler feedback feature to determine the MIP levels and MIP regions that will be sampled for each frame, shading only them in texture space and sampling the shading map in screen space to produce the final image."
 
Sampler feedback - just like the original tiled resources - have always been in a weird space. Even the theoretical benefit is pretty marginal... it's basically cases where you have an anisotropic kernel walk across the edge of a tile boundary and you're unwilling to do borders. Obviously borders are usually completely reasonable for texture data.

For texture-space shading you may want to expend a bit of effort to avoid shading border pixels but it's still kinda in the noise IMO. Without more details on what the "software solution" is and why they need a software *sampling* solution in the second pass (presumably because they are doing indirections per tap instead of borders?), I'm not sure if there's a meaningful way to interpret the results of the test to be honest.
 
It seems to me that texture space shading is about a decade too late. It attempts to cache and reuse shaded texels across multiple future frames. But this assumes that the lighting applied to those texels is constant over those frames. That seems less and less likely with modern games embracing multiple (many) dynamic direct and bounce light sources.
 
UE4 isn't great. Most developers who use their own engine achieve better performance on AMD GPUs.
UE4 is open sourced to AMD (and other IHVs), they can and do submit into it anything which they need to get better performance.
It is in fact a much more fair engine for assessing cross-vendor performance than whatever "most developers" do for AMD based consoles.
 
They do mention, however, that the game has frame pacing issues and that they're worse in DX11
Indeed, but that is an optimization issue, not an API issue, the developer should go ahead and optimize the DX11 path more and the issue should go away, many console games suffered this exact issue at launch, and the developers fixed it later with patches.
 
Back
Top