Why isn't Nvidia's CAS Sharpen a separate setting from NIS in the Control Panel?

Use this program first, execute as admin and disable the mitigations: https://www.grc.com/inspectre.htm and then reboot.

Check if your registry keys are the same as these, if not, change them and reboot again:

Enable Kernel-Managed Memory and disable Meltdown/Spectre patches.
HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Control\Session Manager\Memory Management

FeatureSettings -> 1

FeatureSettingsOverride -> 3

FeatureSettingsOverrideMask -> 3

 

and the original one was broken in the majority of titles.

 

Your cpu doesn't suffer from the mitigations.

 

For the 'AMD vs NVIDIA Drivers: A Brief History and Understanding Scheduling and CPU Overhead' video from 2017 there's an interesting followup discussion thread with clarification / criticism / feedback (see edit in that thread) from game engine / graphics driver engineer perspective:

https://www.reddit.com/r/Amd/comments/6rxun7/why_nvidia_performs_so_much_better_in_dx11/

It's not as simple as hardware scheduling vs software scheduling.

 

It also cites a a quote from an IGNORANT RTG engineer

"The best way to do DX11 is from our GCN Performance tip #31: A dedicated thread solely responsible for making D3D calls is usually the best way to drive the API.Notes: The best way to drive a high number of draw calls in DirectX11 is to dedicate a thread to graphics API calls. This thread’s sole responsibility should be to make DirectX calls; any other types of work should be moved onto other threads (including processing memory buffer contents). This graphics “producer thread” approach allows the feeding of the driver’s “consumer thread” as fast as possible, enabling a high number of API calls to be processed."
​

The best way to do DX11 is using multiple deferred command lists concurrently because you can perform more setup in parallel before playing back the commands to the driver for hardware execution, doing all the commands in a single thread reduces command throughput overall.

PS: We've done and dusted that discussion on this forum repeatedly,

Nvidia removed a DATA HAZARD BLOCK which was both inefficient as it was power guzzling because for graphical tasks, cached states and JIT-Recompiliation can be employed to macro-manage the operations without needing such a block. The hardware retains all other warp and kernel schedulers, and has actually added more with pascal and later.

 

Interesting. I was aware about the latter two, but none of the articles I've referred to even mentioned FeatureSettings "1". It's set to 0 for me. Can you provide a source where this is documented?

 

i've not noticed any significant performance hit on my end really, or nothing that can be noticed with a bare eye.

 

If you have either Ryzen or a CPU past 8th gen from Intel this won't affect you or the hit will be so small you won't even notice anything.

 

Well we know that "a hardware scheduler" which was different between GCN and "Paxwell" GPUs a) has about zero effect on CPU utilization one way or another and b) is actually present in Volta+ GPUs on Nvidia's side. Thus that speculation on HUB part is just pure BS.

HUB themselves have shown many many results where Nvidia GPUs outperform AMD GPUs in DX12/VK in 1080p since then. Curiously they were extremely shy on mentioning how this was supposedly impossible due to a lack of "a hardware scheduler" according to them.

The most likely reason for the CPU load discrepancies between AMD and Nvidia h/w is down to the applications and how they use the respective h/w. In many cases Nvidia has to fix bad or unoptimized code via drivers which leads to higher CPU load. AMD is running into the same issue less often since they enjoy the benefits of being console h/w provider here, meaning that most "big" games are optimized for AMD GPU h/w capabilities by default these days. Still there are cases where DX12/VK renderers run better on Nv h/w so it's not a universal issue and is certainly not something happening because of "a hardware scheduler".

 

I dug this up recently
https://www.reddit.com/r/Amd/comments/6rxun7/why_nvidia_performs_so_much_better_in_dx11/dl9rgax/

 

That's great to know, thanks.

 

Yeah, this is accurate and is the extent of "lack of h/w scheduling" in case of warp execution. Worth noting that AFAIR AMD never even had something like this in the first place, they always did this through s/w stream compiler. So if that is the bespoken "h/w scheduler" then modern AMD h/w is missing it also.

There is another scheduling level which the post mentions and which is more relevant to the topic - the multiple command submission queues. These map to ACEs/HWSes on GCN+ h/w and this may result in some driver side differences between Nv and AMD in CPU loads. But considering that on Nv h/w (pre-Turing at least) all which is happening is serialized submission I don't see how that would result in more than a single percent load difference, likely even less.

 

Out of curiosity I have a 3900X so I recognize those issues won't affect me (at least not nearly as much), but it wouldn't harm performance to enable these registry tweaks correct? I probably won't as the impact from them should be relatively small from what I've read and heard here, but I might test it out for curiosity sake.

I did see on some articles online that they had the featuresettingsoverride and overridemask tweaks present, but I couldn't find anything for the FeatureSettings -> 1 tweak googling around. I didn't look too hard though so perhaps I overlooked something -- so all those 3 tweaks will do is disable the meltdown/spectre patches and enable kernel managed memory? Does kernel managed memory have an advantage over the default?

Sorry to spam you, thanks.

 

Hi, I made those changes based on some registry tweaks for some Windows optimization, with enough time, you can google it all and see what it actually means, I didn't check it much really because I don't care, I just don't want Spectre&Meltdown slowing things down. They are harmless and I always change them on any Windows install, if you are worried about security, don't change them, if any, nothing has ever happened to me. I think this is more on the lines if you execute every single thing you download from any website and you don't even check anything. Remember this is mostly protection against ransomware, which does require you to execute some program that would hijack your computer.

I guess this could help you: https://www.alertlogic.com/blog/a-clear-guide-to-meltdown-and-spectre-patches/

 

Thanks that's helpful