AMD GPUs up to 20-30% faster than Nvidia when paired with low-end CPU

TL;DW Nvidia driver increases CPU load in DX12/Vulkan (and properly multithreaded DX11?) games compared to AMD driver. Results in reduced performance when CPU-bound. Theorized as being due to Nvidia doing software thread scheduling in their driver, versus AMD's hardware scheduling, to improve performance in lightly threaded DX11 games at the cost of increased overhead.

Digital Foundry saw the same behavior when testing Doom Eternal (Vulkan) on an Xbox One X APU. Going from an RTX 2060 to RX 6800 XT at CPU-bound settings doubled framerate.

 

Hardware Unboxed trying to get banned again.

 

This is really only going to be relevant for gaming with low graphics settings at low resolutions. It doesn't really seem relevant unless you are all about competitive FPS and you for some reason are still on an old CPU. I definitely respect the research though.

 

Except that wasn't the case??? With the lower end CPUs, WD:L and H:ZD were limited to less than 100 FPS on average, with minimums closer to 60, on the Nvidia cards. On a high refresh display, those drops are noticeable, not to mention the noticeable frametime stuttering when CPU-bound at low framerates. BTW, the i3-10100 is the same as the i7-6700K and 7700K, which are still relatively popular among gamers who like to stay on CPUs for several generations and just upgrade the GPU.

The Nvidia driver overhead probably explains why I was getting such poor performance in the large modes on CoD Modern Warfare/Warzone and Black Ops Cold War, which are DX12 exclusive, on my previous overclocked i5-8600K with all the settings on lowest. In Cold War especially, it was a nearly constant stuttery 90-100 FPS with maxed out CPU utilization.

 

I'm not saying it's not happening, but if you look at the 1080p ultra and higher numbers the higher tier GPUs pull away convincingly on stronger CPUs, over and above the lesser GPUs on those same better CPU. Your example is a 6c/6t part? That would perform worse than a 4c/8t part in most gaming workloads I would think. No one is pairing that CPU with an RTX 3070+ or RX6800+. I watched the video, I had to rewatch to make sure I didn't misinterpret the data. You can basically see the 5700XT hitting a brick wall in 1080p Ultra on more than one game they test where the 3070/3090/6900 is still scaling fine. Anyone building a desktop system with these higher tier GPUs is going to be using a far better CPU than an i3-10100 or 1600X, the only reason to show this is to make a case for the research. The real world application isn't there.

It was however somewhat interesting to me as someone who uses an eGPU with lesser CPUs, but the bottlenecks for eGPU systems are still elsewhere. It would be interesting to repeat this data again once we start seeing Radeon mobile parts in shipping laptops.

 

4 Hyper-Threaded cores is slower than 6 physical cores.

The 6900 scales fine, the 3090 does not. In WD:L, the 3090 is CPU bottlenecked in all scenarios except at 1440p Ultra with the 5600X.

Again, you're missing point. Desktop gamers generally do not upgrade their CPU/platform as often as they do their GPU. Someone pairing their new RTX 3070 with the enormously popular 3600/3600X is gonna find their performance being no better than the 5600 XT, and much worse than the 5700 XT, both of which should be far slower cards than the 3070. And if they did have a new 5600X, their 3070 is still no better than a 5700XT. That is a problem.

 

I can cherry pick graphs out of the research as well, instead of looking at 1080p Medium in Watch Dogs lets look at 1440p Ultra.

 

Way to miss the point again.

 

Dunno if it helps anyone. I managed to somewhat fix constant micro-stutters or fps drops by reverting to MSFT driver which includes just display driver, control panel and hd audio driver (hdmi) and no Physx and other bloatwares. Had to use wumt x64 to scan my system after DDu'ing 430.xx Dell driver, stock nvidia 440.xx, 460.xx which all had micro-freezes and high battery drain in Optimus mode on 980M with 6700HQ.
Reduced battery drain to from 45W to 7-12W. Also, nvidia GPU utilisation tray icon seems to be buggy and had to un-check it to fix dGPU freezing on windows desktop.

 

So everyone is playing on old CPUs with $550-1500 MSRP (3070 thru 3090/6900) GPUs at 1080p medium settings to realize gains over competing GPUs? I don't get it. If you're not cranking it up to High/Ultra at 1080p (discounting 1440p is the fastest growing segment) with a GPU that expensive you're either a serious competitor in esports FPS titles or ..I can't think of another reason. I crank everything up as high as I can go and only have a 2070 constrained by a 15W 3 year old CPU over a TB connection. I don't know why I would go to medium settings when the FPS gain there is meaningless in all but esports. Every single case where they showed the Radeon outperforming the nvidia cards it was 1080p lowered settings.. for a GPU that can handle better. It's just in the name of finding a thing that most people won't encounter. But if you are in that situation it's good information, never countered that.

I'm not arguing that the research hasn't proven that when wildly CPU constrained (cranking out as many frames as possible at low resolution lower settings) the evidence is that radeon cards/drivers are more efficient, but I am arguing that at the price points of those cards it seems weird to pair it with a super inexpensive CPU when something that would make better use of the cards from either vendor is a fraction of the cost of the video card. Even at the $550-650 price point. Even then a lot of AMD boards can take a better Ryzen CPU without upgrading the motherboard, just drop in a 3800XT, unlike on the Intel side.

Maybe I'm just in the camp of people that think spending another $300 for a CPU when you're spending $750 on a 3080 makes more sense and I am out of touch with the multitudes of people that are using old CPUs with brand new expensive video cards.

 

not everyone games at ultra...the gains are real I say medium is the sweet spot and i personally game on all low with some settings on medium...looks identical to ultra..

but now since i lost my 1060 laptop I game ultra but with geforce now which is like a desktop 2080...streaming is the future so who cares about hardware these days

 

Not everyone plays at ultra settings at 60Hz even in AAA titles. High refresh rates aren’t just for for eSports, besides those games have very low CPU/GPU requirements.

 

Real-world examples...

https://twitter.com/HardwareUnboxed/status/1371028413222318084

https://twitter.com/ANDRSNx2/status/1371100392348606470

 

It's good that they bring this to the limelight so Nvidia can fix it. Imagine how much performance our laptops are leaving on the table.

 

It could probably be improved, but I’m not sure if Nvidia can completely fix it on current GPUs if they lack the hardware scheduling that AMD has.

 

All this even becomes more confusing when you use an eGPU. The RDNA2 GPUs (probably RDNA as well) show horrible performance over TB with any driver newer than April 2020.

https://egpu.io/forums/pc-gaming/amd-egpu-driver-bad-performance-benchmarking/#post-94039

https://egpu.io/forums/builds/2021-...6900-xt-32gbps-m-2-adt-link-r43sg-win10-2004/

In the end, I like the testing and they did a great job. I think their console CPU limitation angle is probably spot on for why AMD took the approach they did. I could definitely see situations where people would benefit from this data and stay away from nVidia GPUs if they are not going to upgrade their CPU. It's super interesting how confusing it made upgrading for those without a certain CPU performance level.

I think it just comes down to what your goals are, what CPU you use and what upgrades you're going to do in the near future if you are only upgrading the GPU. For me nVidia GPUs are a better option due to TB My 2070 completely destroys that 6900XT over TB unless it's using old drivers, even with a BAD CPU limit . My CPU is currently thermal throttling at 10W even undervolted because I need to reapply TIM, and still crushed it. I also tend to crank details up as far as I can go without going below ~40-50fps which favors nVidia. I also play 99.9% single player story based games where you want to crank everything up.

 

Bottleneck with NVIDIA and DirectX 12? MSI GeForce RTX 3090 SUPRIM against MSI Radeon RX 6900XT Gaming X and its own drivers igorslab.de
First of all, I have to put a small paragraph in front of today's post as a preface, because I want to take away the touch of sensation. Nevertheless you have to write about it and you have to test it out emotionlessly beforehand. The video from Hardware Unboxed didn't even surprise me that much, because we were able to make very similar observations on lags and latencies in a current test project (thanks to Fritz Hunter!) And are initially downright desperate at the inconsistency of some measurement data.

 

So the 6800XT/6900XT beating the 3080/3090 at lower resolutions in reviews (usually using top CPUs) was because of driver overhead after all.

 

More testing

 

jesus....do you guys realize what this means for egpu

 

I was waiting for him to address that no one aside from competitive users chasing frames on old cpus are going to hit this as based on the framerates I'm seeing in all these graphs no one is going to use 1080p medium.. just crank to ultra.. Too bad the 1440p result was patreon only. Still good data, but not very relevant even for me with a weaker CPU and eGPU (as Radeon is just HORRID on eGPU for any driver newer than almost 12 months ago)

 

Well, 1080p med seems like a good fit for midrange cards like the 2060 and 5600XT in AC:V if you want to stay above 60 FPS. Interesting that the 5600XT significantly beats the 2060 with all CPUs there.

 

This conclusion (CPU usage) seemed like the best way to simply state it. YouTube timestamps never seem to work in the embeds, it's at 23:36

 

As a former i5-8600K user, I can answer that with an emphatic no. I saw so many games with 100% or close to 100% CPU usage.

 

So you’re planning to buy an AMD GPU next time?

I guess AMD cards suit better for gamers that prefer to keep their CPU & GPU for a long period of time.

AMD users also claimed AMD new drivers always improved their older generation graphic cards performance. I don’t know the truth (data & evidence) behind this statement.

 

Yeah, I think waht they are saying is if you are in that scenario it's better to get a Radeon GPU, as long as you are not also GPU limited I suppose.

 

Depends on how RDNA3 vs. Lovelace shakes out.

 

I was little bit shocked to see mid-range CPUs getting outdated when paired against Turing or Ampere GPUs losing max FPS.

I'm planning to buy AMD CPU+GPU next year. Not much stocks available and most of them are over-priced. Linux support is better on AMD/Intel than nvidia. I lost my mind after upgrading nvidia drivers on Linux and they booted into black screen yet again and I switched to Intel iGPU and disabled nvidia GPU.