Nvidia Has a Driver Overhead Problem, GeForce vs Radeon on Low-End CPUs

I might be totally wrong here, but I'm thinking maybe DX11/DX12 use shared libraries or resources to some extent, and driver is working in a similar way in both API's, but Vulkan is independent so DirectX driver 'optimisations' are not being implemented.. thus not 'in the way' under Vulkan..?
 
demo said:
Pax, Nvidia have taken this approach for DX11 and earlier where there are significant gains, while in DX12 their software approach is a burden because the API handles it. In DX12 it 'gets in the way'.

They probably haven't changed the driver stack completely just for a handful of DX12 games, when DX12 drawcalls are already through the roof by virtue of API.
Click to expand...

DX 12 driver should be quite separate and different from the dx 11 one. But there has to be a good reason why they kept a heavy software stack in dx12 and the software scheduler is probably one of the reasons. But there maybe other ones we dont know of kind of how bob explained it, and would be nice if HU followed up on this with a dev.

New and updated games and engines will be coming on dx12 and would be nice if things didnt get more top heavy over time.

Edit: looks like HU will do a followup on this. Linus looked into it as well on the wan show but thats 2 hours of psychobabble I cant go through without an item menu.

https://twitter.com/HardwareUnboxed/status/1370914994460581888

Replying to
@HardwareUnboxed
I will clarify all this in a video tomorrow. Also the comments from Linus on how to test GPUs were spot on in my opinion. There's really no need to change how GPUs are tested and rather as they said this could be special feature type content.
9:49 PM · Mar 13, 2021·Twitter Web App
Click to expand...
 
Last edited:
HU says they will be doing a lot more testing about this in the next 2 weeks and they pointed to an old video by nerdtechgasm who dealth with this issue some time ago:

[yt]nIoZB-cnjc0[/yt]

He seems well versed.
 
They pretty much nailed it at the end. Anyone still running a 4 or 6 core cpu, and its arguably most people on steam survey, want to consider the overhead going forward when they do gpu upgrades. Itll especially be important for nvidia to consider its next gen and put a hardware scheduler back into its gpus.

Heck a few people I know are considering moving from 4 to mid range 6 cores as an upgrade and saving as much as possible for the gpu whic is stil the best way to upgrade for gaming except now we need to be careful about it.

And its not just pairing a high end gpu with a low end cpu you can easily make the mistake of pairing a mid range gpu like the 3070 and not get the perf you expect vs a more affordable upgrade. This is good info because most people dont consider their gpu upgrades on their aging rigs.

It made sense to do a software scheduler to cool down kepler in the day and overcome software engine limitations that were not thread aware but those days are mostly long gone.
 
I gave you the same answer they did, with draw call benchmarks under each API.

The vast majority of games still benefit from their software scheduling approach. As they said, it's not really an issue for two reasons; first is draw calls are already 10x higher under DX12. Secondly, no one is pairing 3090 with i3. It is something to be aware of, make sure your hardware is suitably matched.
 
demo said:
I gave you the same answer they did, with draw call benchmarks under each API.

The vast majority of games still benefit from their software scheduling approach. As they said, it's not really an issue for two reasons; first is draw calls are already 10x higher under DX12. Secondly, no one is pairing 3090 with i3. It is something to be aware of, make sure your hardware is suitably matched.
Click to expand...

In my case i would have better performance with 5700 than both 3060 or 3070 at 1080p ultra high.I am looking at mins fps.
 
badsykes said:
In my case i would have better performance with 5700 than both 3060 or 3070 at 1080p ultra high.I am looking at mins fps.
Click to expand...


Yeah, when I had my office PC still with an Intel 3777k (a 9 year old CPU) I put my old 2080Ti in there and got abysmal performance in many games.



I certainly would have been better off buying a low or mid range AMD card instead and still playing at lower framerates and minimal details , rather than upgrading the CPU to a 10700F and getting far better performance and quality instead. :rolleyes:
 
demo said:
I gave you the same answer they did, with draw call benchmarks under each API.

The vast majority of games still benefit from their software scheduling approach. As they said, it's not really an issue for two reasons; first is draw calls are already 10x higher under DX12. Secondly, no one is pairing 3090 with i3. It is something to be aware of, make sure your hardware is suitably matched.
Click to expand...

Not really because even in dx11 as HU said most game engines have become thread aware and that explains why the r290 beats a 780ti handily now.

So its actually a problem in dx11 now as well but it wasnt way back when 4-5 years ago when they did that switch.

Its only not a problem if you run a high end cpu 5600x and up.
 
Last edited:
pax said:
Not really because even in dx11 as HU said most game engines have become thread aware and that explains why the r290 beats a 780ti handily now.

So its actually a problem in dx11 now as well but it wasnt way back when 4-5 years ago when they did that switch.

Its only not a problem if you run a high end cpu 5600x and up.
Click to expand...


That's wrong it's only not a problem if you don't hit 100% CPU utilization. Meaning you won't see this limitation on low end CPU's either unless it hits 100% utilization. You're not going to see this really on new games on low end CPU's, but more so older CPU's running newer games. Like the 3770k example, which I personally seen myself when I put my old 2080ti in that system.


This was all explained in HU video.


You're desperately trying to find a problem but when it's really non existent for 99% user base.
 
Exposed said:
Yeah, when I had my office PC still with an Intel 3777k (a 9 year old CPU) I put my old 2080Ti in there and got abysmal performance in many games.



I certainly would have been better off buying a low or mid range AMD card instead and still playing at lower framerates and minimal details , rather than upgrading the CPU to a 10700F and getting far better performance and quality instead. :rolleyes:
Click to expand...

I had a 2080ti paired with a 3770k and it really wasn’t that much worse than the 10900k with a 3080.
 
jimjobob said:
I had a 2080ti paired with a 3770k and it really wasn’t that much worse than the 10900k with a 3080.
Click to expand...


Watch Dogs 2 I was at 100% CPU utilization and around 30-40fps most of the time, even with lowered settings at 1440p. I knew it was the old 3770k CPU holding it back, that part was obvious.



Same 2080Ti, switched to 10700F framerates boosted up to 60-80fps (sometimes even more like 100) at highest detail settings.



Significant performance uplift. I also noticed a significant uplift in almost all VR games, the primary gaming purpose of my office PC aside from work.
 
Exposed said:
That's wrong it's only not a problem if you don't hit 100% CPU utilization. Meaning you won't see this limitation on low end CPU's either unless it hits 100% utilization. You're not going to see this really on new games on low end CPU's, but more so older CPU's running newer games. Like the 3770k example, which I personally seen myself when I put my old 2080ti in that system.


This was all explained in HU video.


You're desperately trying to find a problem but when it's really non existent for 99% user base.
Click to expand...

HU clearly stated otherwise in demo's link. Its been a problem for a while now and we have been seeing 4 core cpus hit 100% since the days of testing Turing. Going forward the driver doesnt seem to be getting more efficient on older cpus.

Most of the testing on this is pretty old but HU said they will be testing a lot more in the next 2 weeks to see where this lies. We'll see.
 
pax said:
HU clearly stated otherwise in demo's link. Its been a problem for a while now and we have been seeing 4 core cpus hit 100% since the days of testing Turing. Going forward the driver doesnt seem to be getting more efficient on older cpus.

Most of the testing on this is pretty old but HU said they will be testing a lot more in the next 2 weeks to see where this lies. We'll see.
Click to expand...


What specifically did they state otherwise? What specific quote/text link? It just seems to me you're not really understanding why Nvidia has this driver overhead, why it/was beneficial, and why the overhead issue won't be a problem for most gamers. I know the usual pitchforks are out but this one is needlessly stretched.
 
The HU video is timestamped and you can see it mentionned at 45min.

HU is doing a cpu scaling vid, that many gpu reviews dont normally do as they tend to review a gpu with the latest and greatest setup.

But the whole point of AIB gpus is to upgrade the gpu on an older rig. Its kind of surprising so few do cpu scaling.

IMO many if not most are still running well under a 5600x cpu so such a review is valid today as game engines are now well thread aware.

You might think its not a problem putting a 3070 with say a 7700k when that setup is running enough fps to your liking but if you are hitting 100% on cpu you will hitch. And you will have less perf than on a radeon gpu or evenb some nvidia gpu that costs a lot less.

So what I think constitutes a problem, IE paying too much for a gpu for a given setup, might not be for you. It gets sujective as to what is a problem or not. IMO as I said I think this is.
 
Last edited:
Exposed said:
Watch Dogs 2 I was at 100% CPU utilization and around 30-40fps most of the time, even with lowered settings at 1440p. I knew it was the old 3770k CPU holding it back, that part was obvious.



Same 2080Ti, switched to 10700F framerates boosted up to 60-80fps (sometimes even more like 100) at highest detail settings.



Significant performance uplift. I also noticed a significant uplift in almost all VR games, the primary gaming purpose of my office PC aside from work.
Click to expand...


Metro exodus max settings 4K no DLSS went from 32 to 37 FPS.
Odyssey went from 54 to 62 fps max settings @4k
Wish I had tried more but the other pc was so old, I had long since stopped benches. Those changes are with both the cpu and gpu being upgraded together.
 
You must log in or register to reply here.
Back
Top