Nvidia Has a Driver Overhead Problem

Like I mentioned in another thread, I am curious to see if NVIDIA has anything to say about this situation.
What extra information could they give us?
Could they potentially expose an option in the driver to disable such scheduling?
What about the "Threaded Optimization toggle on NVCPL? Is it restricted to certain APIs like I have seen some claim or is it global?

Again, I'd like to hear NV about it, not speculation about how the feature works. That type speculation that comes from people who don't have access to driver code or any sort of technical NV only insight.

I wonder, could @ManuelG help us out here? In the sense of "Can you voice our concerns to people inside the company?"

 

I'd like to know why that original thread was closed without an explanation...

 

And I'm wondering why this video was reason enough to post it about 6 times in the forums here already with 3 seperate threads for it added... but hey, I'm glad I'm not a mod to have to worry about such stuff

 

Cause it is a pretty freaking significant discovery. Means that any esports gamer wants to go with an amd gpu.

 

But any 4k gamer wants nvidia gpu becouse amd is bandwidth starved.

 

Yeah, hence why i got the 3090.

But the esports market is significantly bigger than the 4k gaming one. Esports gamers able to gain 20-30 % fps? Not insignificant !

If this knowledge becomes widespread, it could cause a major upset in the gpu market, in amd's favor.

 

lol
wrong.

 

Nvidia fans ware making such a big deal with amd dx11 driver overhead now when tables have turned and nvidia is the one with the issue its not a big deal, its fine actually. Hah, ampere is nvidias vega.

 

Nvidia is marketing 360fps/hz situation though currently. Only unemployed teenagers and mid-rangers should be affected and deserve to know the truth. Pro gamers is the real advertisement. Like e-sport legends were promoting odyssey G9, who the hell would buy such monitor only for MOBAs.

 

That's the ironic part, that when trying to reach 360 fps in esports titles at 1080p, you are more likely to be cpu bottlenecked than gpu, thus would actually get higher fps with an amd gpu.

 

Where? I searched to see if someone already posted it, but didn't find anything.

 

Could you elaborate on that?

I saw some of your comments in the other driver thread, but like I said there I’m not sure I follow your points.

So, you mention this possibly being caused by the NV driver not prerendering, but in Vulkan/DX12 I’d read that the developer controls prerendering and you cannot force a different value via the driver. If I’m incorrect in this, let me know/apologies, but that was my understanding. If it’s Dev controlled it would be the same for NV and AMD no?

Your other point I recall was regarding aggressive power management iirc. As for power savings, to me, this sort of differential in performance if it is power related would indicate Nvidia’s power algorithm is being too aggressive and leaving too much performance on the table which would still be a problem, no?

That NerdTech video I linked earlier in this thread goes over the differences in the NV and AMD schedulers and seems to explain why exactly this differential occurs where if I’m understanding right NV’s is better suited to lightly threaded games, but can cause these issues in games that are multithreaded well already.

Crysis 3 would be a very interesting use case to test as its both DX11 and multithreaded quite well according to Digital Foundry.

 

I attempted to post about it in the NV driver section since I didn’t see it posted already — it turns out I didn’t look far enough back in the history for yours and the other one had been removed already. It makes sense mine was removed, but I’m not sure why the original got cut. Ah well, at least we have this one, thanks for posting it correctly @RealNC

 

Theres a lot more than just FlipQueue when it comes to the pre-preparation of tasks to be sent to the graphics card,
You cannot directly target a specific value with d3d12 and vulkan, but the setting does affect peak performance to some degree.

Waiting on nvidia to finish their investigation, make any changes and do an Andrew B article.

Theres misinformation about "Schedulers" being spread around, the Turing and Ampere parts are pseudo hardware schedulers that are driver interactive.

eSports gamers are playing games built with D3D9 and 11 which historically operate better on nvidia.

and no, i don't consider fortnut an esport.

Theres some factors involved with the power saving that also exists to prevent dropping to and getting stuck at low power 3D, the power management used on Geforce cards is Util based focused around the GPU itself.

The entire chip being awake and at high enough clocks under low resolution/setting gaming could see util go below the point where demotion occurs (roughly 40-45%), because Vram usually is the first part to demote to lower clocks this can see games stuck at 100% vram controller usage and an inability to clock back up - prefer max is a workaround here but nvidia doesn't actively go through finding what games hit this possible case and making the default prefer max.

instead the gpu can turn off parts of the chip so that low details are registering a higher util of the chip for what is active, but effectively perform slower.

the HWU demonstration of higher cpu util is also explainable as using wait/busyloops in the UMD to keep things that need synchronisation in line, AMD has done this in the past too. These are not actually doing anything with the cpu apart from artificially raising the cpu util (usually they also do not contribute extra heat either).

 

Well not surprised, dx12 was AMD thing by default.. mantle.
DX11 is nvidia thing, from Turing on dx12+ was designed to be more flexible with hw async, but still not enough, I guess?

This resizable BAR could change some stuff, they should test with that too and compare for "driver overhead".

 

So on that topic too..
Changing the threaded optimization on/off in NVCP may change something or not?

 

not relevant to d3d12 or vulkan afaik.

 

Reason why Nvidia has much better DX11 performance is that their driver is much more efficient than AMD's at managing high-level graphics workloads.

When you get into working with a low-level API such as DX12 or Vulkan, the application relies less on driver efficiency and more reliant to how the application has been coded by developers to take advantage of the hardware itself. Much more specific options are available to the developer to manage the GPU that would be otherwise managed by the driver when using DX11, but the developer needs to be more careful of what is coded in their application to take proper advantage of each model of hardware.

It's not so much that Nvidia has more driver overhead, but AMD had a less-efficient driver to handle high-level work. Now that DX12/Vulkan allows AMD to fully stretch its hardware without relying much on driver optimization, the API performance between Nvidia and AMD is closer than in the past.

 

Really interesting stuff, thanks! Yup, will be interesting to see more discussion after further investigation.

 

This isn't anything new... NVidia drivers not only require faster CPU but also faster CPU frequency on first core because it's mainly used for this stuff(ie. it's not multithreaded correctly), which meant that Intel was better than AMD(CPUs) for NVidia cards. DX9 was purely first core, DX11 was still mainly first core, only DX12 is apparently supposed to have multithreaded overhead and it's AMD(CPU/GPU combos) benefiting most.

But the real question is, is it fixable programmatically? Could driver changes help? How much of it is DirectX issue?