Nvidia's 3000 Series RTX GPU [3050, 3060, 3070, 3080, 3090 now with TIs]

@sir doris If you go back to the BFV video with the old i7, the guy “fixed” his performance regression with the 1660 ti by overclocking his ram from 1866 to 2133 and also improving primary and secondary timings. Maybe he just saved enough clock cycles by lowering latency to give his gpu more room to run, or maybe nvidia drivers themselves are actually sensitive to cache and memory latency.

 

@sir doris Igor also doesn’t measure cpu usage in any way, and instead tries to infer performance. He also makes a guess that it has something to do with how nvidia drivers manage asynchronous compute, but I’m not really clear on how he got there without actually doing some kind of profiling. I’m not sure bfv dx11 can do asynchronous compute, but the issue is visible there.

 

A Far Cry Primal comparison between FuryX and 980Ti on the other end of spectrum. 4k-res with high settings and low-fps gpu bound settings( <50fps on both ) with FuryX having higher fps and lower CPU usage at the same time.

 

@gamervivek The more I look for tests the more I find ones that show AMD using less cpu for about the same performance, even if just slightly.

Spoiler

This Far Cry 5 is using an 8700K with a 1080 and Vega64. Vega is a little bit faster and uses just a tiny bit less cpu in this test.


Here's Rage 2 running a little bit better on AMD with about the same CPU usage. I guess this is the avalanche engine? Whatever the Just Cause team uses.

 

In fairness to Igor he admits he doesn't really have the time atm to investigate this further with the detail that's required, so he is basically just throwing out suppositions at this point.

It is interesting though that he's observing actual noticeable artifacts as a result of this (albeit in a manufactured situation). Again not hard data which is what we want, but at least in his brief experience, this is producing behaviors which actually affect the presentation beyond just dropped frames and less CPU headroom:

 

It's a little bit better in overall fps in Rage, but what stood out to me in particular was the consistency of the rendering graph in ms - the Radeon is noticeably more stable. No massive drops on the Geforce, but overall a far more jagged line.

 

In other news, NVIDIA broke their own Ethereum-limiter on GeForce RTX 3060 with the new GeForce 470.05 beta-drivers.
The drivers are available from NVIDIAs developer site (requiring registration) and apparently from Windows Update for Windows Insider users.

https://www.computerbase.de/2021-03/mining-bremse-geforce-rtx-3060-umgehbar-eth-hashrate/
https://www.hardwareluxx.de/index.p...e-der-geforce-rtx-3060-umgangen-2-update.html
https://pc.watch.impress.co.jp/docs/news/yajiuma/1312085.html
https://videocardz.com/newz/pc-watch-geforce-rtx-3060-ethereum-mining-restrictions-have-been-broken

 

Not sure if that wasn't an intentional change:

 

I recently played through Control on a GTX 1080 in DirectX 12 mode because the level load times were much faster there. Rendering performance didn't seem noticeably different in either mode.

Though I've read that the texture LOD changes might be faster in DirectX 11. In DirectX 12 you see low resolution textures being swapped quite often.

 

Intentional or not, we have fresh drivers without the limiter. The limiter being active on other cards if there's more than one has been already bypassed.

 

The texture pop-in definitely was lessened on my gtx1660 in DX11 compared to 12, but there's actually this mod which fixes the texture pop-in completely in that mode:

https://community.pcgamingwiki.com/files/file/2035-control-blurry-textures-fix/

I never see texture pop-in with it installed, even when running from a HDD. Even less than on a PS5.

DX12 though has always been a disaster for me in this regard, this mod does nothing for it.

 

PCGH tests CPU limited performance.

They conclude that when focusing on low level API games, AMD is noticeably faster.

 

Three Vulkan games there show a different result though - 3090 is faster than 6900XT.
That's 3 out of 11 of "low level API games". NV is generally faster in DX11 too.
Frametime graphs aren't that different either, with AC Valhalla being the only game noticeably better on 6900XT.
Again it's not clear what is the issue here or if it even is CPU related and not API related or app specific.

 

Would have been nice if they specified which api each of the games was using. I have no idea what api planet zoo uses, for example.

 

I do agree that its clear we are seeing symptoms of something, but the tests have not yet been sufficient to lead to a conclusion of the cause quite yet.
We are still very much still in the diagnosis stage, hope to see more tests and information released.

 

I would still like to see a straightforward cpu utilization plot with equalized performance (in game frame limiter). It should be pretty easy to see which uses more cpu time.

Could easily be something that’s only a consideration in dx12, for example.

 

Needs to be more than that I think. We should think on how to profile the CPU time per frame to see where it is being spent on different systems and in different games.

 

I have zero expectations of any reviewers touching a profiler, but profiling drivers may actually be difficult. Windows Performance Toolkit, maybe?

 

It was visible since the first games. DX12 on nVidia hardware doesnt work right. For example Hitman 1 with DX12 is slower on my 4C/8T and 2060 notebook than DX11 in a CPU limited scenario. There is a software(API) overhead involved to get nVidia GPUs running. And without proper multi-threading the nVidia DX11 driver is just superior to DX12...

 

It wasn't intentional, Nvidia have confirmed a developer left the unlocked code in a beta driver which they've since removed.