Discussion [HWUB] Nvidia has a driver overhead problem. . .

[Hitman928]

HWUB did an investigation into GPU performance with lower end CPUs for gaming and published their results. In the comments they mention they tested more games than they showed in the video, all with the same results. As we all (should) know, in DX11 AMD had the low end CPU performance issue, but it looks like this has flipped for DX12/Vulkan. HWUB mentions they think their findings hold true for DX11 as well, but as far as I can tell, they only tested DX12/Vulkan titles so I don't think they have the data to backup that statement and doubt it is true.

(342) Nvidia Has a Driver Overhead Problem, GeForce vs Radeon on Low-End CPUs - YouTube

 

[Panino Manino]

I'm not able to watch it right now. Could you summarize?

They gave more data.
Just to "confirm what we already knew and others are testifying".

 

[Mopetar]

The name of the video gave me the impression that they'd uncovered some new or important details.

 

[Hitman928]

The name of the video gave me the impression that they'd uncovered some new or important details.

From a very brief skip through the video, it seems like in this video they tried to match 'realistic' configurations (e.g. mid range cpu + mid range GPU) and showed that there is still an issue for Nvidia with these configurations.

 

[Mopetar]

From a very brief skip through the video, it seems like in this video they tried to match 'realistic' configurations (e.g. mid range cpu + mid range GPU) and showed that there is still an issue for Nvidia with these configurations.

I guess then it's kind of what was concluded here, but it's good to have testing and verification of this.

At least a new CPU is a lot easier to get than a new GPU that will have these problems.

 

[DAPUNISHER]

the type of people who can afford to buy a high quality 4K display are the same people who can afford a high-end GPU to render a native resolution and image on it.

This hypothesis neglects that many of us use a PC instead of a console to couch game on inexpensive 4k TVs.

 

[Mopetar]

This hypothesis neglects that many of us use a PC instead of a console to couch game on inexpensive 4k TVs.

Upscale 1080p to 4K then. It's basically what the consoles do. However, even if you have a less powerful GPU, the inexpensive 4k TV likely caps out at a 60 Hz refresh rate, which a mid-range GPU can actually hit in a number of titles.

If you look at the TPU reviews for the 3060 which has an MSRP of $330 (yeah I know that's a pointless number right now) they still average right around 50 FPS at 4K across the 23 games that they benchmark and only a small handful of titles drop below 30 FPS, which could probably be alleviated by turning down the settings.

Unless you have a newer Nvidia GPU (or a newer AMD GPU that gets support for their DLSS equivalent) you wouldn't get the capability anyways. Even the bargain basement 4K televisions cost around the same price as a mid-range GPU, or theoretically would if miners were paying two or even three times MSRP for practically any card.

 

[Leeea]

I'm not able to watch it right now. Could you summarize?

More data, and even more shocking results. With my cpu, in some tests even an old rx580 was keeping up with the GeForce product line.

examples:
Fortnight @ 1080p /w i7-4790k (my cpu):
the rx580 outperforms the entire GeForce product line.
fortnight ceased to receive performance gains after the Vega 56.


Hitman2 @ 1080p /w i7-4790k:
Vega 56 faster then the entire GeForce product line
no meaningful performance improvements for rx6000 over vega56


Horizon Zero Dawn @ 1080p /w 4790k:
rx6000 > 3080 > vega56 > 3060
rx6000 has a meaningful frame rate boost over vega56 ( +20 min fps )
rx580 & gtx1060 unplayable


Red dead redemption & Rainbow Six Siege ( 4790k @ 1080p ) the difference in framerates between vega56, rx6000, rtx3000 were not meaningful to me.


Watch Dogs Legion ( 4790k @ 1080p ):
rx6000 @ 64 fps min
vega 56 @ 60 fps min
rx580 > 3060ti (min fps)
rx580 & all GeForce < 60 fps min


-------------------------------

What is shocking to me is an rtx 3080 would have likely been a side grade verses my vega56, specifically in the min fps numbers. I was originally leaning toward the rtx for its raytracing and dlss. Ultimately, my bot just worked better on AMDs website, and that was all there was to it.


Upgrading to a rx6000 series was still a considerable upgrade for me. Especially at 1440p* and 4k**. However, considerable performance remains to be unlocked with a future cpu upgrade.

*my monitor
**my TV, which I am currently using for my Sekiro play through. My Vega56 was unable to do 4k /w Sekiro with med/low settings, however the rx6900xt can easily do Sekiro 4k with max settings.

If you look at the TPU reviews for the 3060 which has an MSRP of $330 (yeah I know that's a pointless number right now) they still average right around 50 FPS at 4K across the 23 games that they benchmark and only a small handful of titles drop below 30 FPS, which could probably be alleviated by turning down the settings.

For me this would not be true. All of the games I play on the couch are controller / reaction time games, and 60 fps is required. The TV is only capable of 60hz /w black frame insertion.


My old Vega 56 averaged 50 fps-ish with Sekiro @ 4k with the settings turned down. Sekiro is a game focused on rhythm, timing, and reaction time. It is internally capped and designed for 60 fps. I found the game difficult to play at 4k - 50 fps. I turned it down to 1440p - 60 fps and upscaled, but the upscale just was not as crisp and I struggled in the game. I ended up quitting the game for a while.


With my new card I can play 4k @ 60 Hz, and it makes the entire game just easier and more enjoyable. I have made more progress in the last week then I made in the entire time I owned it previous. The finer detail really seems to make a world of difference.

 

[Timorous]

Upscale 1080p to 4K then. It's basically what the consoles do. However, even if you have a less powerful GPU, the inexpensive 4k TV likely caps out at a 60 Hz refresh rate, which a mid-range GPU can actually hit in a number of titles.

If you look at the TPU reviews for the 3060 which has an MSRP of $330 (yeah I know that's a pointless number right now) they still average right around 50 FPS at 4K across the 23 games that they benchmark and only a small handful of titles drop below 30 FPS, which could probably be alleviated by turning down the settings.

Unless you have a newer Nvidia GPU (or a newer AMD GPU that gets support for their DLSS equivalent) you wouldn't get the capability anyways. Even the bargain basement 4K televisions cost around the same price as a mid-range GPU, or theoretically would if miners were paying two or even three times MSRP for practically any card.

If the 3080 cannot even maintain 60 fps minimums in some games at 1080p medium it does not matter how GPU bound you make the settings you are not going to get smooth 60fps+ gameplay.

 

[Mopetar]

If the 3080 cannot even maintain 60 fps minimums in some games at 1080p medium it does not matter how GPU bound you make the settings you are not going to get smooth 60fps+ gameplay.

That has to be one hell of a CPU bottleneck. Either that or there's a spiritual successor to Crysis I'm not aware of.

 

[Timorous]

That has to be one hell of a CPU bottleneck. Either that or there's a spiritual successor to Crysis I'm not aware of.

4790k in watch dogs legion and given the 9400F was 51 fps min in CP2077 I expect it would be the same there too.

It was also sub 60 fps minimums with the R5 1400 in AC Valhalla, CP2077, Horizon Zero Dawn, Shadow of the Tomb Raider and Watch Dogs Legion. In all of those titles the 1400 + 5600XT was just as good or better at those settings. The 3080 would only offer an improvement if you wanted to run higher settings but your frame rate would not budge. This applies down the stack to the 3060Ti so if you can get the 6700XT or 3060Ti at MSRP (ha) and have an older CPU thr 6700XT will improve frame rates more than a 3060Ti will. Heck if you are coming from a 5600XT the 3060Ti might be a regression in FPS *edit* if you want 60fps minimums.

 

[Mopetar]

The R5 1400 results aren't terribly surprising (or at least not to me) based on the data I had posted from GameGPU benchmarks. I think people forget that while the original Zen was a pretty impressive accomplishment from AMD, it was still clearly behind Intel. The only area it really clearly won was offering a high core count on top of better SMT for some truly incredible performance for anything that could rake advantage of all of the threads.

Otherwise the 1400 performance was pretty similar to what you could get out of a 2600k. I don't think it's too surprising to expect that a 10-year old CPU would bottleneck games. People just tend to forget that AMD had a lot of catching up to do (which they ultimately did) so some of their older CPUs are a bit "older" than their year of release would suggest.

 

[DAPUNISHER]

so some of their older CPUs are a bit "older" than their year of release would suggest.

Yeah, no. When Zen released it was up against Intel 7 series. Zen has aged much better in the respective price ranges. Weaker IPC or not. And It is silly to compare it to a 2600k for a number of reasons, but most importantly, price at release. Back to the timeline - The 7700K 4/8 for $350 on release, is considered the worst CPU purchase of 2017. The i3 and i5 have aged even worse. Particularly given how Intel rushed the 8 series out so quickly, in response to Zen. Anyone that bought 7 series got seriously boned.

The 7600K was $242, a Ryzen1400 was $170, and even the faster 1500x was under $200. I know which I'd rather be playing the latest demanding titles on. Especially given how resource heavy the latest iteration of 10 is, and with apps like discord running. And the win10 lite meme ain't going to save that badly aging 4/4 disappointment either. That instead of testing the 7 series, he tested Haswell, is because so little changed in that period. Now THAT is what I call being older than the release date suggests.

 

[Mopetar]

I'm making the comparison against the 2600k because that's about where it landed in terms of gaming performance at the time of launch. Generally the original Zen CPUs weren't as good in terms of either IPC or clocks when they were matched up against the Intel equivalents of the time. Where they did shine was being able to get an 8C/16T CPU for under $1,000 and in being able to run highly threaded applications quite well because their SMT was better than Intel's.

From the tables I posted on the first page, the 7700K won't bottleneck in the newest AC game, whereas the 1400 shouldn't be paired with anything newer than a GTX 1080 for that particular title. The table on the second page shows a 7700K is good enough if paired with a 6900XT that it won't be CPU bound in Forza Horizon 4. If you bought a 7700K for gaming, it's still holding up well enough.

You can add price in there and make arguments related to that, but that becomes a bit more subjective when arguments about value enter into the mix. But the GameGPU data suggests that going from a 4770K to a 7700K makes a big difference, at least for the titles that I compiled the data for. It gives a 27.8% uplift for a 6900XT and a 31.3% uplift for a 3090 in Forza. The 3090 gets a 26.6% bump in AC as by going from a 4770K to a 7700K. I think the reason the 4-series was tested it because the 7-series doesn't bottleneck as nearly as much.

 

[DAPUNISHER]

I'm making the comparison against the 2600k because that's about where it landed in terms of gaming performance at the time of launch. Generally the original Zen CPUs weren't as good in terms of either IPC or clocks when they were matched up against the Intel equivalents of the time. Where they did shine was being able to get an 8C/16T CPU for under $1,000 and in being able to run highly threaded applications quite well because their SMT was better than Intel's.

From the tables I posted on the first page, the 7700K won't bottleneck in the newest AC game, whereas the 1400 shouldn't be paired with anything newer than a GTX 1080 for that particular title. The table on the second page shows a 7700K is good enough if paired with a 6900XT that it won't be CPU bound in Forza Horizon 4. If you bought a 7700K for gaming, it's still holding up well enough.

You can add price in there and make arguments related to that, but that becomes a bit more subjective when arguments about value enter into the mix. But the GameGPU data suggests that going from a 4770K to a 7700K makes a big difference, at least for the titles that I compiled the data for. It gives a 27.8% uplift for a 6900XT and a 31.3% uplift for a 3090 in Forza. The 3090 gets a 26.6% bump in AC as by going from a 4770K to a 7700K. I think the reason the 4-series was tested it because the 7-series doesn't bottleneck as nearly as much.

This is one of those replies, where I feel it muddies the waters too much. The entire premise of this thread, is the overhead with Nvidia cards. Then you site examples of a 1400 w/Nv and Intel w/AMD? I confess you lost me. There may have been a great point in there, but I am too dumb to understand it.

What I do understand: there is nothing subjective about CPU pricing in the context I presented it. Fact: the 1400 launched as a direct competitor in price to the top i3 and locked i5 of its time. Those are the correct CPUs to compare it too, full stop. The i7 OTOH, was always a high end $325-$375 CPU each gen, until AMD gave them a cup check. And value may be subjective, but who at this stage in history, would assert a 7700K on a dead socket, is a better value than 1st gen AM4?

The Kaby Lake aged like milk. Steve hits the bullseye with this article https://www.techspot.com/article/1523-worst-cpu-gpu-purchases-2017/

I should mention that article was written 11/13/2017. It has only gotten worse since then.

 

[Mopetar]

This is one of those replies, where I feel it muddies the waters too much. The entire premise of this thread, is the overhead with Nvidia cards. Then you site examples of a 1400 w/Nv and Intel w/AMD? I confess you lost me. There may have been a great point in there, but I am too dumb to understand it.

It's just some data that I'd posted earlier in the thread that someone else linked from a review. Since the charts are dynamically generated I copied the data into a table and did some basic statistics. The conclusion is that both Nvidia and AMD GPUs will see bottlenecks on older CPUs, but Nvidia more so than AMD does. First generation Zen with low core counts bottlenecks extremely hard because of lower clock speed and IPC on top of this. Anything older than Skylake for Intel will also cause issues for both to some degree.

The point I was making was that based on the data, a 7700K won't cause the same types of problems (though it's probably getting close to the point where it will) you'd see from a Haswell CPU and the GameGPU data shows an uplift of around 30% from going from a 4770K to a 7700K so I strongly disagree with your contention that it didn't matter if Steve tested with either of those since the results would be the same. That's the main point I was making.

Even Steve's article is more about the value being bad because less than a year later Intel actually started competing with what AMD was offering now that they couldn't just keep charting $350 for an unlocked 4-core CPU. However that doesn't mean that it's bad because it will now create bottlenecks in a lot of games for high-end GPUs. From the data I've looked at, that isn't the case at all.

For AC:Valhalla specifically (you can see the data for yourself if you scroll down to the CPU test in the GameGPU data) not even the Ryzen 2000 series chips are enough to avoid bottlenecking a 3090 or 6900XT. Forza Horizon 4 shows pretty much the same where you need at least a Zen 2 CPU not to bottleneck a 6900XT, but with a 7700K you're still GPU limited.

 

[IntelUser2000]

So the spotlight on Nvidia having driver overhead problems are being shone upon when AMD is making a comeback in graphics.

Testing on a 2600K and Core i3 7100 CPU, the Polaris 10 based cards benefit immensely with 8GB VRAM on Shadow of the Tomb Raider only because the 8GB VRAM allows DX12 to be used without spilling out of VRAM and DX12 under that scenario is seriously faster than DX11.

Polaris 10 with 8GB VRAM DX12 is like twice as fast as 8GB VRAM DX11. GTX 970 that otherwise performs quite close can't compete at all when the game supports DX12.

AMD cards allow similar/better performance with cheaper setups than Nvidia. So they are not only competitive on the high-end, they allow you to enjoy with older setups.

 

[TheELF]

Testing on a 2600K and Core i3 7100 CPU, the Polaris 10 based cards benefit immensely with 8GB VRAM on Shadow of the Tomb Raider only because the 8GB VRAM allows DX12 to be used without spilling out of VRAM and DX12 under that scenario is seriously faster than DX11.

If this is the case then how valid are these claims?
The game benchmarks can fit inside the 8Gb and never have to swap out, but in actual gameplay how realistic is it for a game not to have to swap out, and even if it wouldn't have to swap out doesn't the different mem structure of the consoles force the games to stream in data constantly anyway?

 

[Justinus]

If this is the case then how valid are these claims?
The game benchmarks can fit inside the 8Gb and never have to swap out, but in actual gameplay how realistic is it for a game not to have to swap out, and even if it wouldn't have to swap out doesn't the different mem structure of the consoles force the games to stream in data constantly anyway?

VRAM has nothing to do with it, look at RTX 3090 benchmarks, sometimes beaten by a 5600xt or 580 with 1/3 or 1/4 of the VRAM when paired with a older CPU.

There is nothing about DX12 that necessitates the same game requires more memory. That's not how it works.

 

[JoeRambo]

Frankly R1400 was and is horrible CPU. People forget that it is has 2 CCX'es active in a 2+2 setup. Cores have access to just 4MB of L3 maximum.

That means inter thread communications have to cross CCX boundary and on ZEN1 that was especially slow affair. Any workload that overloads intercore comms is gonna suffer.
AMD has advantage here, cause they are parents of this abomination and probably force critical driver threads on same CCX as game render threads, so there is less intercore overhead and they continue to scale somewhat.

Compare to proper quad core with 6MB of L3:

 

[Racan]

Well Digital Foundry did talk a little about this issue but it seems to me that they are throwing shade at Hardware Unboxed's testing and are trying to minimize AMDs advantage here.

 

[AtenRa]

Well Digital Foundry did talk a little about this issue but it seems to me that they are throwing shade at Hardware Unboxed's testing and are trying to minimize AMDs advantage here.

Heh not a big issue for many they say ??

According to Steam Hardware Survey almost 42% have a Quad Core CPU like the old 4C 4T Core i5s and old 4C 8T Core i7s. There will also be some newer 4C 4T Core i3s and a few of the latest 4C 8T Intel gen 10 Core i3s, plus the older ZEN/+ cpus like the Ryzen 1400/1500 etc.
Also there is almost a 29% that have a 6-core and from those the vast majority will be those early Ryzen 1600/X-2600/X and 3600/X. From Intel there will include Gen 8 8700/k and Gen 9 Core i5s like the 9600/k and Gen 10 10400/f/10600/k etc.

So according to Steam, more than half of the users could be affected by this.

Steam Hardware & Software Survey

In the majority of the latest DX-12 games all those CPUs above (Both AMD and Intel) display some sort of driver overhead when paired with NVIDIA graphics cards especially with Turing and Ampere.
Add that in a few months NVIDIA will release Lower-End Ampere cards like the RTX3050Ti that will have performance close to current RTX2060/Super and we can begin to see that many users will try to update form GTX1060 to latest sub 300$ NVIDIA Ampere cards and lots and lots of them will still be using those older CPUs. The result will be that lots of them will not be able to get the performance they will see in the reviews simple because of that CPU overhead.

The situation would be wayyyyyy worse if GPU prices where close to MSRP, because lots of those users would buy even faster Ampere GPUs than what they will buy today and the CPU Overhead problem would bee even worse for them as those high-end Ampere GPUs have the same problem even at 1440p.

I strongly believe that people with older CPUs should take this problem seriously when they are going to upgrade their GPUs next time and more reviewers should look in to this and provide more data to inform the community.

 

[Mopetar]

It really depends on the game, but it's becoming more common for games to make use of more than just four cores, so I think we're clearly at a point where anyone building a gaming rig shouldn't consider less than six or maybe even eight just to be safe.

If you look at the charts the original Zen CPUs with 6 or more cores tend to do a fair bit better than the quad-core products, especially in games that seem to be using the CPU to do more sophisticated enemy AI or other more computationally heavy tasks.

 

[DiogoDX]

Well Digital Foundry did talk a little about this issue but it seems to me that they are throwing shade at Hardware Unboxed's testing and are trying to minimize AMDs advantage here.

RT on AMD in CP2077 is out now. You can bet that for this they will have a video with all the details on how Nvidia is faster.

 

[Head1985]

RT on AMD in CP2077 is out now. You can bet that for this they will have a video with all the details on how Nvidia is faster.

Ofcourse.From now i will just ignore all NVfoundry ehm "digital foundry" GPU reviews.They are biggest nv shills out there.
Btw dx12 cpu driver overhead is pretty much reversal situation from dx11 era.All new AAA games are dx12/vulkan now and AMD have HUUUGE advantage here.AMD now have 20-30% more performance in all CPU bottleneck scenarios under dx12/vulkan.

 

[JoeRambo]

Ofcourse.From now i will just ignore all NVfoundry ehm "digital foundry" GPU reviews.They are biggest nv shills out there.

Really? I've found their card reviews spot on. They have obviuos bias on rendering technology advances, so things like RT and DLSS techniques are worth a premium to them and it shows in their reviews. For example:

AMD Radeon RX 6900 XT review: the Digital Foundry verdict

Digital Foundry review the AMD RX 6900 XT - the ultimate BIG NAVI. Game benchmarks, Smart Access Memory testing, ray tracing performance and more.

www.eurogamer.net

Hard to disagree with their take. Heck with the hindsight, AMD Super Resolution is not out yet and even if it was released today, how many games would support it? Meanwhile Nvidia is pushing UE4 plugins that enables DLSS support without much work.