Hardware & Technical Modified/stripped down (mostly GPU) drivers and related tangents

Starting with 552.22 my "Base Profile.nip" will now force DLSS to use profile "E" which has significant motion clarity advantages over the defaults in most apps. However, this will require a recent version of the DLSS libraries to be used.

 

Vulkan beta released the other day, modified as above. Wasn't going to post them as they usually release a new Game Ready WHQL driver within a few days, but that hasn't shown up yet, so may not arrive until Tuesday.
 
New NVIDIA driver fixes some encoder issues and is probably worth a look.

I've also started including three different versions in my modified driver package. Standard version is my usual "any close processor" floating affinity for GPU driver interrupts, while the CPU_03 and CPU_12 lock affinity to the specified logical core for potentially lower latency and more consistency. CPU_03 should be compatible with pretty much any vaguely modern CPU, while CPU_12 is intended for AMD parts with one or more full eight-core CCXes (no six or twelve core CPU variants) and Intel parts with at least eight P-cores (9th Gen and newer i7s and i9s) with SMT/HT enabled (it locks affinity to the first logical CPU of the seventh physical core, which tends to be the least contested on these parts, without having to cross any light latency Fabric or ring/mesh boundaries). In any case, the difference will be extremely minor, probably imperceptible to most, so don't expect miracles or anything.


Additionally, I've updated my Intel 2.5G Ethernet (i225) and my MediaTek MT7922/AMD RZ616 Wifi6E drivers to the latest versions, if anyone has a use for them. I've tuned the interrupt affinities, stripped out some unneeded components (the IHV router), exposed some more settings, and set the defaults to minimize latency at the expense of other parameters. These drivers are for Windows 10 and Server 2016/2019/2022 only and are not signed, so need to be installed with the procedure mentioned in the first post.

 

Attachments

  • Intel-i225_1.1.4.43_mod.zip
    288.4 KB · Views: 76
Note that if one is using an AMD part with CPPC preferred cores enabled (which is going to include pretty much every AM5 processor) there may be additional considerations to optimal interrupt affinity. Preferred cores likes to condense light and moderately threaded tasks on to a few logical CPUs. Static interrupt affinities should be placed on less preferred cores.

If you have an AM5 CPU, you'll want to open HWiNFO's sensor page and look at the CPPC preferred core order shown here:
hfNeE9r.png


If the first number I've underlined is a 1, 2, or 3, do not use the CPU_03 package. If the second number I've underlined is a 1, 2, or 3, do not use the CPU_12 package.
 
Reworked my NVIDIA 560.70 package. I pulled the core-specific interrupt affinity versions as this should really be tuned to the specific system.

I also applied the new NVENC encoder session limit patch and manually integrated the newest Vulkan 1.3.290 runtimes.

 

The combined package contains the 24.10.1 drivers (in the "Display" folder) for recent AMD GPUs and 24.9.1 ("Display2") for older, but still supported models. Keep in mind the procedure in the first post if one wants to install these; they are neither signed nor include an installer.
 
New 566.36 drivers:

Note the installer file name has a "Test" at the end of it because I had to manually do some tweaking to get NvCleanstall to work with this driver version, which is the first to default to the new NVIDIA App (which I did not include...standard control panel only here). I tested the installer and it works as expected, but left the file name as I had already zipped it up.
 
AMD 25.1.1 beta drivers, stripped down as noted in the OP:

NVIDIA 571.96 drivers (slightly newer than the RTX 5000 series press release driver, from the CUDA 12.8 Toolkit), some tools, and DLSS4 binaries from the new Cyberpunk patch:

I did not include any customized profile this time, so I recommend manually disabling CUDA P2 state and enabling ReBAR globally with the included nvidiaProfileInspector.

If you want to use the new DLSS4 transformer model ray reconstruction, you'll need to replace the target game's DLSS .dll files with the included ones and use ProfileInspector to force DLSS profile "J" for that game. Results will vary.
 
AMD 25.1.1 beta drivers, stripped down as noted in the OP:

NVIDIA 571.96 drivers (slightly newer than the RTX 5000 series press release driver, from the CUDA 12.8 Toolkit), some tools, and DLSS4 binaries from the new Cyberpunk patch:

I did not include any customized profile this time, so I recommend manually disabling CUDA P2 state and enabling ReBAR globally with the included nvidiaProfileInspector.

If you want to use the new DLSS4 transformer model ray reconstruction, you'll need to replace the target game's DLSS .dll files with the included ones and use ProfileInspector to force DLSS profile "J" for that game. Results will vary.
As far as I understand DLSS4 will support all RTX cards (starting with 2060). Can we appeal to Elite developers to introduce DLSS4 support in it? I think a lot of people would vote in favor of it.
Or is it a very difficult task for Cobra?
I think FSR is not so good on NVIDIA cards. And in this case a lot of work is put on the CPU's shoulders.
 
Or is it a very difficult task for Cobra?

If the version of Cobra ED is built with exposed the data needed for temporal upscaling, we'd surely have it by now.

I think FSR is not so good on NVIDIA cards. And in this case a lot of work is put on the CPU's shoulders.

FSR 1.0 is mediocre at best, and needs a well anti-aliased input to not be crap, which ED can't provide. FSR 1.0 is what we've got because it's purely spatial and doesn't need the temporal information TAA, FSR 2+, XeSS, or DLSS require.
 
Back
Top Bottom