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

[Twinz]

What about windows 10 hardware accelerated GPU scheduling under advanced settings in game mode?

What hardware is it using when this is enabled?

I tried it yesterday and my FPS and frame times looked good on the FPSVR hud but it *felt* like I was getting micro-stuttering to the point I had to change it back.

 

[oguzsoso]

What about windows 10 hardware accelerated GPU scheduling under advanced settings in game mode?

What hardware is it using when this is enabled?

I tried it yesterday and my FPS and frame times looked good on the FPSVR hud but it *felt* like I was getting micro-stuttering to the point I had to change it back.

it has been said that scheduling in question is different

i tried and enabled hardware scheduling and saw no improvement in cpu bound benchmarks

What about windows 10 hardware accelerated GPU scheduling under advanced settings in game mode?

What hardware is it using when this is enabled?

I tried it yesterday and my FPS and frame times looked good on the FPSVR hud but it *felt* like I was getting micro-stuttering to the point I had to change it back.

i did some benchmarks with it on and off and found no difference in performance in cpu bound situations

i also felt like it was actually making system somewhat more clunky. i don't even know what it was supposed to bring as an advantage... so i disabled it as well. since it provides zero documentable benefits for an end user like me.

 

[Puppetmaster]

I'd like to see comparisons playing at 1080p High / max and 1440p. Most people with decent cards are not playing at medium.

Might investigate some games to test my 5930k and 3080 with at 1440 to see how much I'm losing.

 

[Nexus18]

I'd like to see comparisons playing at 1080p High / max and 1440p. Most people with decent cards are not playing at medium.

Might investigate some games to test my 5930k and 3080 with at 1440 to see how much I'm losing.

Yup exactly, I get that they are wanting to show the issue is there but it was utterly daft showing only 1080P and medium settings Might as well have just dropped to 720p.

Most people I know with 2080s/2080ti/5700xt/3070/6700 and above cards are at 1440p/4k and running max/high settings.

 

[andybird123]

So essentially, it's only a serious problem if:

- you play at 1080P (bit pointless having a 3080 etc. at this res.)
- run less than max/ultra/very high settings (seems a bit odd to do this if you have the likes of a 3080....)

People who are playing at 1440p and/or 4k with max settings need not worry too much.... I skipped through the video and couldn't see any tests at 1440P or higher with max settings???

Guess my 2600 @ 4.1 will do until amd get their act together and match intel for price to performance but I play most games at 4k 60 now anyway so most of the work is on the 3080 and it can maintain a locked 60 @4k here now except for cyberpunk 2077



Still though, nvidia clearly have a problem here, which needs addressed.

You also need to be running a CPU that's 6+ years old, at stock settings so no overclocking (whilst hanging about on a forum called "overclockers") or a CPU that was essentially a laughing stock when it was released.

 

[humbug]

I'd like to see comparisons playing at 1080p High / max and 1440p. Most people with decent cards are not playing at medium.

Might investigate some games to test my 5930k and 3080 with at 1440 to see how much I'm losing.

Right, everything needs to be compacted down to the performance level of your CPU. No one needs to see the progression difference between higher end / newer CPU's, that's just rude.

 

[oguzsoso]

https://youtu.be/3O-ePSXS5Yw?t=526

so, in these scenes a theoritical equivalent amd gpu with equivalent ray tracing performance would bottleneck %20 less, no?

it seems like this video is very outdated. 5600x later patched to utilize SMT. i will try to find a similar video post-smt fix patch.

Found a vid with 1.12 version. CPU usage is still low but dunno. It should be smt fix applied version. maybe game doesnt use that much thread, here you go

https://youtu.be/UGp8WuleqEg?t=44

Still bottlenecks below 60 fps.

Honestly, even if i had a 5600x, after learning about this truth, i would still stay away from Nvidia GPUs.

a 5600x should not drop below 60 fps in any game, ever. and yes, cyberpunk is bad with its optimization. but with an amd gpu, it would drop %20 less frames, so...

 

[humbug]

https://youtu.be/3O-ePSXS5Yw?t=526

so, in these scenes a theoritical equivalent amd gpu with equivalent ray tracing performance would bottleneck %20 less, no?

it's clear that 5600x cant handle rtx pyscho's bhv structure load when paired with a 3080. 3080 chills at %75-80 gpu usage, while 5600x struggles to feed it and frames drop below 50, even to 43s, and cannot go to 60 at all. it's clear that 3080 is capable, it's the cpu that holds it back (notice that this is at 1440p dlss balanced with very ultra settings. if you go 3080 dlss quality, then you get gpu bound 50 fps. but the problem is, you have the balanced alternative which can gain you gpu bound 60+ fps, but cpu limits it)

even 5600x cannot be declared "immune" to this problem, to be honest

the situation in general is really bad. needs at least some slight improvements, a %10-15 would be good...

Its a shame they only have the overall CPU utilisation usage, we can't see what individual threads are doing, the GPU is bottlenecked but its not bad at between 80% and 99% utilisation.

This game is just really badly optimised, DX12 or not if its not well coded it can still choke on a single worker thread and that's what appears to be what's going on here, tho we cannot know as we can't see what the individual threads are doing, but the CPU has 50% left in the tank.

This is an Nvidia sponsored title, Nvidia are known to hurt thier own performance in the games they sponsor if it hurts the performance of their competition more and the one thing that AMD's Hardware Scheduler doesn't deal with as well with as Nvidia is a scenario where the bottleneck is with a single worker thread.

 

[Nexus18]

Honestly, even if i had a 5600x, after learning about this truth, i would still stay away from Nvidia GPUs.

a 5600x should not drop below 60 fps in any game, ever. and yes, cyberpunk is bad with its optimization. but with an amd gpu, it would drop %20 less frames, so...

Here's a good video showing a 3080 with i5 10600K and a 6800xt with a 5600x, max settings @ 1440p (including assassins creed and cyberpunk) No DLSS used either so it's about as like for like as you can get.

Doesn't seem like there's any issues here, FPS are much the same and where it should be for a 3080 vs 6800xt.

 

[humbug]

This is the battle between them, as has been said in a really rather good video someone posted in this thread some pages back.

Its in Nvidia's interest for the game to use as few Draw Call threads as possible, their Software Scheduler can distribute the load across other available threads, but its at a cost of about 20% of the CPU's cycles. this is where you just have the software that sits between the Game and the hardware doing all the work.

AMD's Scheduler is on the GPU, the game connects directly to that Scheduler, there is no dedicated worker thread, the Scheduler connects a task directly to an available thread, if there are 4 threads on the CPU the GPU has 4 options to hook individual tasks to, if there are 16 threads it can connect 16 tasks to 16 threads.

So AMD want games that do as much work on the CPU as possible, Because Nvidia need 20% of the CPU just to function, so the more Draw Calls the game has the more CPU resources the game needs and with that will bottleneck sooner on Nvidia's GPU's, that is what we are seeing.

 

[humbug]

Here's a good video showing a 3080 with i5 10600K and a 6800xt with a 5600x, max settings @ 1440p (including assassins creed and cyberpunk) No DLSS used either so it's about as like for like as you can get.

Doesn't seem like there's any issues here, FPS are much the same and where it should be for a 3080 vs 6800xt.

Because there isn't, you look at the GPU loads on both sides its very high, the CPU or the driver overhead doesn't come into it here, this is simply a performance comparison between the RTX 3080 and RX 6800XT, nothing else is at play here

 

[Nexus18]

Because there isn't, you look at the GPU loads on both sides its very high, the CPU or the driver overhead doesn't come into it here, this is simply a performance comparison between the RTX 3080 and RX 6800XT, nothing else is at play here

Yup I know, it was in response to this:

a 5600x should not drop below 60 fps in any game, ever. and yes, cyberpunk is bad with its optimization. but with an amd gpu, it would drop %20 less frames, so...

That and as said, cyberpunk is just a poorly optimised game.

 

[oguzsoso]

Its a shame they only have the overall CPU utilisation usage, we can't see what individual threads are doing, the GPU is bottlenecked but its not bad at between 80% and 99% utilisation.

This game is just really badly optimised, DX12 or not if its not well coded it can still choke on a single worker thread and that's what appears to be what's going on here, tho we cannot know as we can't see what the individual threads are doing, but the CPU has 50% left in the tank.

This is an Nvidia sponsored title, Nvidia are known to hurt thier own performance in the games they sponsor if it hurts the performance of their competition more and the one thing that AMD's Hardware Scheduler doesn't deal with as well with as Nvidia is a scenario where the bottleneck is with a single worker thread.

yeah but according to these findings, a 2600 similarly bottlenecks everything. it has plenty of cores and threads.

Here's a good video showing a 3080 with i5 10600K and a 6800xt with a 5600x, max settings @ 1440p (including assassins creed and cyberpunk) No DLSS used either so it's about as like for like as you can get.

Doesn't seem like there's any issues here, FPS are much the same and where it should be for a 3080 vs 6800xt.

heh... this is pure rasterization test. rtx needs a great deal of cpu power.

Spoiler: a

in that video, without rtx, both cpus seem to feed the gpus just fine

i'm investigating situations with rtx enabled-cpu bound situations

rtx adds an approximate %25 more cpu overhead in this game, proof

with rtx disabled, dlss ultra, completely cpu bound, we get 83 fps

with rtx ultra, dlss ultra, we get 65 fps

both situations are completely CPU-bound

this means that rtx on, besides its load on GPU, adds a hefty %25 cpu overhead.

in that video, this is non-existent. in that video i posted, it drops to 45 fps, right? without rt overhead, it would render 56-60 frames (and even then, the area i posted is more cpu-heavy, compared to the place in testing games' video


video i shared has a lot more cars, npcs and in a more crowded place. video you shared is at a more chill place where there are less traffic and npcs, so it's even less cpu-bound

my 5600x video,

yours,

it's virtually impossible with a 5600x to get pure, locked 60 fps in cyberpunk when paired with a nvidia gpu with rtx enabled. you have to make compromises, lower crowd density, lower cascaded shadow range and such.

we don't know if it will be possible with an amd gpu, because there's no amd gpu that can match 3080's ray tracing power, yet. maybe it will be possible, and then maybe this problem will get more exposition. im trying to pre-expose it but

1) cyberpunk needs amd rt support
2) amd cpu tests must be done with a 6900xt at 720p, due to match the 3080's rt performance at higher resolutions

if 6900xt with rt can stay above 60 fps in a situation where 3080 drops with 5600x, then it's clear that overhead affects this aspect of the game as well.

and lastly, let's no assume more games won't be like cyberpunk. developers love releasing unoptimized games. this is to their benefit. they will do the minimum work, and gamers will buy newer and more powerful cpus to accomodate and compensate.

 

[humbug]

@oguzsoso the 2600 does, its per core performance is significantly lower than a 10600K and 5600X.

To use the most extreme but very real example, the FX 8350 has 8 cores, but its beaten by an Intel CPU of the same era with just 4 cores and 4 threads. In the same way a 20 thread 10900K is often match by a 16 thread 5800X in high thread workloads today, each core is just that much faster.

This is the battle between them, as has been said in a really rather good video someone posted in this thread some pages back.

Its in Nvidia's interest for the game to use as few Draw Call threads as possible, their Software Scheduler can distribute the load across other available threads, but its at a cost of about 20% of the CPU's cycles. this is where you just have the software that sits between the Game and the hardware doing all the work.

AMD's Scheduler is on the GPU, the game connects directly to that Scheduler, there is no dedicated worker thread, the Scheduler connects a task directly to an available thread, if there are 4 threads on the CPU the GPU has 4 options to hook individual tasks to, if there are 16 threads it can connect 16 tasks to 16 threads.

So AMD want games that do as much work on the CPU as possible, Because Nvidia need 20% of the CPU just to function, so the more Draw Calls the game has the more CPU resources the game needs and with that will bottleneck sooner on Nvidia's GPU's, that is what we are seeing.

There is another aspect to this, AMD know they can put more cores on a CPU than Intel can, they are quite happy to see a 5950X run along in games at 90%, Intel aren't. AMD / Nvidia / Intel all have their motives, 16 cores on a mainstream 'for gamers' platform may seem strangely overkill to most, not to AMD and not to Intel, for them to come you have to build it first.

 

[oguzsoso]

@oguzsoso the 2600 does, its per core performance is significantly lower than a 10600K and 5600X.

To use the most extreme but very real example, the FX 8350 has 8 cores, but its beaten by an Intel CPU of the same era with just 4 cores and 4 threads. In the same was a 20 thread 10900K is often match by a 16 thread 5800X in high thread workloads today, each core is just that much faster.



There is another aspect to this, AMD know they can put more cores on a CPU than Intel can, they are quite happy to see a 5950X run along in games at 90%, Intel aren't. AMD / Nvidia / Intel all have their motives, 16 cores on a mainstream 'for gamers' platform may seem strangely overkill to most, not to AMD and not to Intel, for them to come you have to build it first.

yet you said cyberpunk bottlenecking could be due to single thread worker bottleneck, that's why i gave that answer

despite being single thread bottlenecked, a 2600 can provide %20 more fps with an amd card in situations where it's not maxed out. that's what i'm saying

what if the overhead happens within the rendering thread? why assume that nvidia's driver puts this load across empty cores/threads_? do you have any documentation for that? (i hope you dont get this in an offansive tone, i'm asking this to get more knowledgable, and i always respect the information)

from what i see, 2600 gets less fps with an nvidia card even in the situations where it's not maxed out. if nvidia scheduler really used empty cores/threads, it shouldn't have rendered more fps with an amd card, no?

clearly, per-core performance can still be affected by overhead, if that's case, in which, 5600x bottlenecking the 3080 below 60 fps still makes my statement valid, it bottlenecked more because of per-core performance lacked (of course due to being horrendously optimized)

 

[humbug]

yet you said cyberpunk bottlenecking could be due to single thread worker bottleneck, that's why i gave that answer

despite being single thread bottlenecked, a 2600 can provide %20 more fps with an amd card in situations where it's not maxed out. that's what i'm saying

what if the overhead happens within the rendering thread? why assume that nvidia's driver puts this load across empty cores/threads_? do you have any documentation for that? (i hope you dont get this in an offansive tone, i'm asking this to get more knowledgable, and i always respect the information)

from what i see, 2600 gets less fps with an nvidia card even in the situations where it's not maxed out. if nvidia scheduler really used empty cores/threads, it shouldn't have rendered more fps with an amd card, no?

clearly, per-core performance can still be affected by overhead, if that's case, in which, 5600x bottlenecking the 3080 below 60 fps still makes my statement valid, it bottlenecked more because of per-core performance lacked (of course due to being horrendously optimized)

You're right, i'm guessing Cyberpunk is just using that very old fashioned approach of just dumping everything into a single thread, at least with Nvidia which doesn't have a Scheduler unless its passed through Nvidia's software layer first, so task A, B, C, D all dumped on the same thread, AMD's GPU Scheduler says task A to thread 1, task B to thread 2, task C to thread 3...

 

[oguzsoso]

You're right, i'm guessing Cyberpunk is just using that very old fashioned approach of just dumping everything into a single thread, at least with Nvidia which doesn't have a Scheduler unless its passed through Nvidia's software layer first, so task A, B, C, D all dumped on the same thread, AMD's GPU Scheduler says task A to thread 1, task B to thread 2, task C to thread 3...

Hmm, so potentially there can still be improvements

Lots of ryzen 2600 friends over the years were shocked when they couldn't get locked 144 fps in their LoL games. tbh, if this issue affected them, then it's a shame, because their cpu practically worked at funny %15-25 usages

 

[humbug]

Hmm, so potentially there can still be improvements, nvidia at least managed to leverage these threads, then?

Lots of ryzen 2600 friends over the years was shocked when they couldn't get locked 144 fps in their LoL games. tbh, if this issue affected them, then it's a shame, because their cpu practically worked at funny %15-25 usages

Could be, Cyberpunk has been in development for more than a decade, its a very old engine, 10 years ago most games just crammed everything into one thread, both Nvidia and AMD in their own ways have successful forced the change to what we see today, its certainly possible that more work could be done with Nvidia to improve thread scheduling.

 

[oguzsoso]

https://imgsli.com/NDY2MzI

Check this out, a health %17 increase with a 2700x.

https://youtu.be/3O-ePSXS5Yw?t=526

so, in these scenes a theoritical equivalent amd gpu with equivalent ray tracing performance would bottleneck %20 less, no?

it seems like this video is very outdated. 5600x later patched to utilize SMT. i will try to find a similar video post-smt fix patch.

Found a vid with 1.12 version. CPU usage is still low but dunno. It should be smt fix applied version. maybe game doesnt use that much thread, here you go

https://youtu.be/UGp8WuleqEg?t=44

Still bottlenecks below 60 fps.

Honestly, even if i had a 5600x, after learning about this truth, i would still stay away from Nvidia GPUs.

a 5600x should not drop below 60 fps in any game, ever. and yes, cyberpunk is bad with its optimization. but with an amd gpu, it would drop %20 less frames, so...

"

• Enabled Ray Tracing on AMD graphics cards. Latest GPU drivers are required."

ok, rt is coming to cyberpunk. i hope someone can test cpu overhead with ray tracing enabled on cpu bound configs (720p high rtx settings) with a 3070 and 6800xt at equivalent settings

 

[Grim5]

This is the battle between them, as has been said in a really rather good video someone posted in this thread some pages back.

Its in Nvidia's interest for the game to use as few Draw Call threads as possible, their Software Scheduler can distribute the load across other available threads, but its at a cost of about 20% of the CPU's cycles. this is where you just have the software that sits between the Game and the hardware doing all the work.

AMD's Scheduler is on the GPU, the game connects directly to that Scheduler, there is no dedicated worker thread, the Scheduler connects a task directly to an available thread, if there are 4 threads on the CPU the GPU has 4 options to hook individual tasks to, if there are 16 threads it can connect 16 tasks to 16 threads.

So AMD want games that do as much work on the CPU as possible, Because Nvidia need 20% of the CPU just to function, so the more Draw Calls the game has the more CPU resources the game needs and with that will bottleneck sooner on Nvidia's GPU's, that is what we are seeing.

good thing I went for the 16 core then hehe, 6 core is outdated in 2021 consoles showing the way, 8 core = minimum for serious gamer