GPU power draw dropping with Raytracing turned on

[Oropher3]

When running on my RTX 2080 Super i can run it raytraced (partially) mostly ok with drops to the low 40s.
But some area's will get me in the mid 20s.
When i checked i saw my GPU temperature dropping while the load remained at 99 %.
The power draw though dropped significantly.
Switching off raytracing immediately returns the power draw to normal.
I would think this actually an internal bottleneck of the 2080 Super not having enough RT cores to accompany the rest of the rendering process.
Can someone confirm or deny this theory?

 

[motox73]

How does the CPU look like with RT on and off ?

 

[Oropher3]

Well, as the framerate drops CPU usage drops and there is no CPU bottleneck.
I have a 3800x which shouldn't pose a problem.
GPU usage is at 99 % but it somehow isn't using the full power of the GPU.
And i think that is because of a lack of RT cores.




There are some area's in this game where i do get a bottleneck, but there it seems as if it is RAM-speed related, as CPU usage then doesn't go above 30 % but GPU usage still drops, ah well, but that has nothing to do with my original post.

 

[SigilFey]

1.) Are you using any overclocking? If so, disable that and see if the power weirdness continues. The OC software may not know how to deal with the RT tech.

2.) It could be a simple matter of the GPU not needing any additional power. If the processor is already at capacity rendering the RT data, drawing more power won't matter. There's nothing it can do with it.

3.) We really don't want to be running any processor at 99%. Ideally, we want things to be below capacity at all times while still performing smoothly. When a GPU or CPU is running at full capacity for extended periods, we would call that "bringing the system to its knees". This is done sparingly for purposes of burn-in or benchmarking, but it will drastically decrease the lifespan of the hardware if done regularly for long stretches of time.

In general, it does sound a little funny that the GPU, alone, should be at capacity...but the voltage is simultaneously dropping off. My guess is that it's a software issue somewhere. Maybe it's just drivers. Perhaps it's the monitoring software, itself. You could try running GPU-Z instead. That's the one I go to. It's very reliable and detailed.

 

[Oropher3]

1. no OC in place.
2. Could be
3. Well, normally the GPU is at 98 %, but i always considered that good, knowing there is no bottleneck.
4. I don't think it's a software issue, as this behaviour Cyperbunk has showed for me, from the start.
It just seems like the level of RT being rendered is too huge for the 2080 Super.
I use MSI afterburner, but i wouldn't know if it's possible it reads out the voltages wrongly, but i am actually sure the data is correct.
As while this happened the temperature of the GPU also dropped. When i disabled RT, usage went down from 99 to 98 % and temperatures went up about 5 degrees.
Anyway, i also think it happens after longer periods of play, where it kinda looks like the game may be overloading some pipeline of the GPU due to not efficiently removing some of the raytracing data.

 

[SigilFey]

3. Well, normally the GPU is at 98 %, but i always considered that good, knowing there is no bottleneck.

When i disabled RT, usage went down from 99 to 98 % and temperatures went up about 5 degrees.

This concerns me. That should not be the case. In fact, that would indicate the GPU is the bottleneck, but there is absolutely no way that a 2080 Super should be anywhere near 90% at all times...unless you're trying to run at 4K, Ultra, at unlimited FPS. Even still, I don't think it would live in the 90% range. Something's wrong there.

Anyway, i also think it happens after longer periods of play, where it kinda looks like the game may be overloading some pipeline of the GPU due to not efficiently removing some of the raytracing data.

If this is true, it's likely a heat issue within the case. Technically...it could be an issue with drivers, but this is starting to sound more like the monitoring software is not working correctly or you have a physical issue with that card. I wouldn't panic or anything, but it may be well worth the money to have that looked at by a good gaming PC shop, if you can find one in your area. Giving them a few screenshots of the behavior would definitely help.

Let me expound a bit here to explain my reasoning:
What should be happening is the following, and I'm going to assume that you're running 1080p-2K with settings High or Ultra. First, we should be seeing lots of fluctuation in GPU usage, voltage, and heat depending on the scene in the game. Something like V's apartment or traveling the wider expanses of the Badlands should probably be shifting between 60%-70% GPU useage. We might see usage in the 80%-range for situations like viewing the whole city skyline from a near distance (like the corporate park area). Occasionally, the GPU may spike into the 90%-range for really intense scenes, like gunfights in detailed areas with lots of smoke and stuff on screen. Voltage and temps should shift around accordingly. Also, of special note, for a game like this, CPU usage should be fairly independent of GPU usage. The CPU should likely spike during loading of new areas while driving around or have little peaks when in combat with a lot of enemies at once. That should not affect the GPU at all really, though you may see hitching or stuttering for brief periods if the CPU is trying to catch up. Very, very rare to see a CPU hit over 70%-80%, though, as many of it's cores will simply not be used by most game's functions.

Now, I don't own an RT card, so I'm crossing into territory that I can only relate to vestigially, but I've got a sense of what to expect based on what I have read and learned. RT is new tech, so there will be "funny" things that happen. I've seen lots of experimentation that showed surprisingly good performance from RT scenes, while performance was actually worse using rasterization. (Legend of the Tomb Raider is notorious for such scenes.) This is not altogether unexpected, because while the tech involved in RT is overall more complex than rasterization/transport methods, one of its major selling points for the future is that it will negate the need for lots of individual systems to handle various lighting elements (like glow maps, diffusion, colored surface highlights, and especially reflections like chrome, mirrors, or glass.) Thus, it's not inherently worrying to find scenes here and there that actually perform better on RT...but that should absolutely not affect overall performance.

Thus, for you to have that GPU, being driven consistently to that level, on top of seeing the weird voltage issues, along with what sounds like universal heat and performance degredation when switching to rasterization...

...that's a LOT of weirdness all at once. Pretty good indicator that, in professional parlance, "Sumpin' ain't right, yo."

 

[Thraksor]

It's pretty simple why you'll see a Turing GPU do what you described:

The game's normal pre-render > render pipeline (raster pipeline, or final output to a display) is waiting on results from the RT cores (and/or maybe Tensor cores if DLSS is involved). It's generally always the RT cores as they are expensive and relatively few for what we are asking of them and are doing calculations that make anything before them look so very simple.

The problem is basically the result of a bottleneck within the GPU's own architecture such that using RT is very much capable of becoming a bottleneck in terms of raw latency.

Let's just say that your 2080S and it's memory bandwidth can render the raster image you're asking of it (settings and resolution) in 8.33ms (120 FPS). Well, when the RT cores and the tensor cores get involved, you've introduced two new sources of bottlenecking to the equation. Can the RT cores also do their job in 8.33ms? Can the NGX code for DLSS being pushed to the Tensor cores do their job in 8.33ms? The answer is almost universally no, at least for the RT cores. They might easily be lagging behind at as much as 50ms (20 FPS) for higher resolution RT operations. NGX is generally less of an issue as most of the grunt work for that has already been done on a supercomputer and your GPU is just replaying what that supercomputer found like an audio or video track. It's the same way Unreal Engine's pre-compiled lighting is cheap to run if you don't take into account the weeks it might have taken a map author to cook the lighting into the map so that your CPU and GPU just need to "replay" that lighting data rather than calculate it from scratch.

So anyway, when a situation arises where the RT cores become the bottleneck as I described above, your GPU can only run as fast as it's weakest link and the result is that the raster portion (the traditional CUDA cores) are basically sitting idle and waiting 41.67ms *every frame* for those weak and relatively slow RT cores to do what they're going to do. And that is because the frame can not be buffered or output to a display until the CUDA cores get the data from the RT and Tensor cores. It's like how a City Bus could haul ass down the highway if it didn't need to stop every so often. That latency will kill it's overall speed.

The power draw dropping is because your CUDA cores are at rest a massive amount of time relative to their real ability/VRM potential while they are waiting an inordinate amount of time on the RT cores to calculate what is supposed to go where a traditional raster shader would have gone. Usage is still going to remain high because it is essentially looking at all the GPU's cores and VRM draw in total and not just the CUDA cores anymore, from Turing forward.

That brings us to another interesting fact about Turing's architecture: it was never designed to run all the CUDA, RT, and Tensor cores all at once. Nvidia built them from the ground up with some pretty gnarly power limitations. You can't use all your CUDA cores AND all your RT cores AND all your Tensor cores at once. RT operations sap something like 18% or more from the raster engine of the GPUs. Tensor cores take about 2-4%. That is one of the big reasons that Ampere GPUs require so much more wattage: Nvidia addressed the weakness inherent in all Turing GPUs by allowing all three types of cores to have their own dedicated MOSFETs and ICs within the VRM architecture in the reference boards. And we still saw AIBs beefing those up in anticipation of it not being enough and that's how you get GPUs like the EVGA RTX 3080 FTW3 Ultra or ASUS Tuf 3080 which packed their boards with so much extra power draw capabilities. My 3080 FTW3 Ultra with the XOC BIOS from EVGA will draw 420W and that is insane, but that's where all of this is going if we want to have real RT capabilities without limits in the PC market: 850W minimum and 1000-1200W systems as a median. And those are single GPU and not SLI or CFX GPU builds (those are kind of already dead tech now).

So even in the best of conditions, roughly 20% of the power available to a Turing GPU's raster cores is lost by enabling RT and DLSS. And 20% power loss means essentially... Nvidia couldn't just break all their GPU draw rates and GPU stability when they have a billion different combinations of power and frequency rates as we have in the AIB GPU market, so their trick to maintain GPU stability was to just have sections of CUDA transistors that can be shut off to let the VRM's MOSFETs and IC focus on either or both RT and Tensor cores first with perhaps as much as 20% or maybe even more of the normal CUDA cores simply being shut off in order to power those optional sections of the GPUs and that came at the cost of raw FPS whenever RT or NGX are enabled.

When that happens, ultimately, the draw rate of the card is going to fluctuate like crazy while the GPU is still going to report full usage. It's just not reporting exactly what is being used anymore, at least not with the kind of accuracy we expected in Pascal and previous generations. Usage is almost a misnomer now because it means nothing without context for whether you're using RT or DLSS and on what specific Turing+ GPUs.

 

[northwold]

This concerns me. That should not be the case. In fact, that would indicate the GPU is the bottleneck, but there is absolutely no way that a 2080 Super should be anywhere near 90% at all times...unless you're trying to run at 4K, Ultra, at unlimited FPS. Even still, I don't think it would live in the 90% range. Something's wrong there.



If this is true, it's likely a heat issue within the case. Technically...it could be an issue with drivers, but this is starting to sound more like the monitoring software is not working correctly or you have a physical issue with that card. I wouldn't panic or anything, but it may be well worth the money to have that looked at by a good gaming PC shop, if you can find one in your area. Giving them a few screenshots of the behavior would definitely help.

Let me expound a bit here to explain my reasoning:
What should be happening is the following, and I'm going to assume that you're running 1080p-2K with settings High or Ultra. First, we should be seeing lots of fluctuation in GPU usage, voltage, and heat depending on the scene in the game. Something like V's apartment or traveling the wider expanses of the Badlands should probably be shifting between 60%-70% GPU useage. We might see usage in the 80%-range for situations like viewing the whole city skyline from a near distance (like the corporate park area). Occasionally, the GPU may spike into the 90%-range for really intense scenes, like gunfights in detailed areas with lots of smoke and stuff on screen. Voltage and temps should shift around accordingly. Also, of special note, for a game like this, CPU usage should be fairly independent of GPU usage. The CPU should likely spike during loading of new areas while driving around or have little peaks when in combat with a lot of enemies at once. That should not affect the GPU at all really, though you may see hitching or stuttering for brief periods if the CPU is trying to catch up. Very, very rare to see a CPU hit over 70%-80%, though, as many of it's cores will simply not be used by most game's functions.

Now, I don't own an RT card, so I'm crossing into territory that I can only relate to vestigially, but I've got a sense of what to expect based on what I have read and learned. RT is new tech, so there will be "funny" things that happen. I've seen lots of experimentation that showed surprisingly good performance from RT scenes, while performance was actually worse using rasterization. (Legend of the Tomb Raider is notorious for such scenes.) This is not altogether unexpected, because while the tech involved in RT is overall more complex than rasterization/transport methods, one of its major selling points for the future is that it will negate the need for lots of individual systems to handle various lighting elements (like glow maps, diffusion, colored surface highlights, and especially reflections like chrome, mirrors, or glass.) Thus, it's not inherently worrying to find scenes here and there that actually perform better on RT...but that should absolutely not affect overall performance.

Thus, for you to have that GPU, being driven consistently to that level, on top of seeing the weird voltage issues, along with what sounds like universal heat and performance degredation when switching to rasterization...

...that's a LOT of weirdness all at once. Pretty good indicator that, in professional parlance, "Sumpin' ain't right, yo."

Just a note / clarification: from what I've heard, at least, it's actually fairly normal for a GPU to run at 100% or close to it in gaming, especially if there's no limit on frame rate. If it's running 100% when idle then there is a problem.

If GPU isn't hitting or approaching 100% with a maxed out game that is not so old it can't use all the GPU's capabilities, my understanding of it is that that indicates a bottleneck elsewhere in the system.

 

[Oropher3]

It's pretty simple why you'll see a Turing GPU do what you described:

The game's normal pre-render > render pipeline (raster pipeline, or final output to a display) is waiting on results from the RT cores (and/or maybe Tensor cores if DLSS is involved). It's generally always the RT cores as they are expensive and relatively few for what we are asking of them and are doing calculations that make anything before them look so very simple.

The problem is basically the result of a bottleneck within the GPU's own architecture such that using RT is very much capable of becoming a bottleneck in terms of raw latency.

Let's just say that your 2080S and it's memory bandwidth can render the raster image you're asking of it (settings and resolution) in 8.33ms (120 FPS). Well, when the RT cores and the tensor cores get involved, you've introduced two new sources of bottlenecking to the equation. Can the RT cores also do their job in 8.33ms? Can the NGX code for DLSS being pushed to the Tensor cores do their job in 8.33ms? The answer is almost universally no, at least for the RT cores. They might easily be lagging behind at as much as 50ms (20 FPS) for higher resolution RT operations. NGX is generally less of an issue as most of the grunt work for that has already been done on a supercomputer and your GPU is just replaying what that supercomputer found like an audio or video track. It's the same way Unreal Engine's pre-compiled lighting is cheap to run if you don't take into account the weeks it might have taken a map author to cook the lighting into the map so that your CPU and GPU just need to "replay" that lighting data rather than calculate it from scratch.

So anyway, when a situation arises where the RT cores become the bottleneck as I described above, your GPU can only run as fast as it's weakest link and the result is that the raster portion (the traditional CUDA cores) are basically sitting idle and waiting 41.67ms *every frame* for those weak and relatively slow RT cores to do what they're going to do. And that is because the frame can not be buffered or output to a display until the CUDA cores get the data from the RT and Tensor cores. It's like how a City Bus could haul ass down the highway if it didn't need to stop every so often. That latency will kill it's overall speed.

The power draw dropping is because your CUDA cores are at rest a massive amount of time relative to their real ability/VRM potential while they are waiting an inordinate amount of time on the RT cores to calculate what is supposed to go where a traditional raster shader would have gone. Usage is still going to remain high because it is essentially looking at all the GPU's cores and VRM draw in total and not just the CUDA cores anymore, from Turing forward.

That brings us to another interesting fact about Turing's architecture: it was never designed to run all the CUDA, RT, and Tensor cores all at once. Nvidia built them from the ground up with some pretty gnarly power limitations. You can't use all your CUDA cores AND all your RT cores AND all your Tensor cores at once. RT operations sap something like 18% or more from the raster engine of the GPUs. Tensor cores take about 2-4%. That is one of the big reasons that Ampere GPUs require so much more wattage: Nvidia addressed the weakness inherent in all Turing GPUs by allowing all three types of cores to have their own dedicated MOSFETs and ICs within the VRM architecture in the reference boards. And we still saw AIBs beefing those up in anticipation of it not being enough and that's how you get GPUs like the EVGA RTX 3080 FTW3 Ultra or ASUS Tuf 3080 which packed their boards with so much extra power draw capabilities. My 3080 FTW3 Ultra with the XOC BIOS from EVGA will draw 420W and that is insane, but that's where all of this is going if we want to have real RT capabilities without limits in the PC market: 850W minimum and 1000-1200W systems as a median. And those are single GPU and not SLI or CFX GPU builds (those are kind of already dead tech now).

So even in the best of conditions, roughly 20% of the power available to a Turing GPU's raster cores is lost by enabling RT and DLSS. And 20% power loss means essentially... Nvidia couldn't just break all their GPU draw rates and GPU stability when they have a billion different combinations of power and frequency rates as we have in the AIB GPU market, so their trick to maintain GPU stability was to just have sections of CUDA transistors that can be shut off to let the VRM's MOSFETs and IC focus on either or both RT and Tensor cores first with perhaps as much as 20% or maybe even more of the normal CUDA cores simply being shut off in order to power those optional sections of the GPUs and that came at the cost of raw FPS whenever RT or NGX are enabled.

When that happens, ultimately, the draw rate of the card is going to fluctuate like crazy while the GPU is still going to report full usage. It's just not reporting exactly what is being used anymore, at least not with the kind of accuracy we expected in Pascal and previous generations. Usage is almost a misnomer now because it means nothing without context for whether you're using RT or DLSS and on what specific Turing+ GPUs.

Thanks for the replies,
I know that it is indeed normal for a GPU to run at 99 % if you want unlimited framerate.
Thraksor really gives the right explanation here, BUT, i am still wondering why i can run the same places of nightcity at 40-50 fps afer just loading the game (with raytracing) and those same scenes can drop below 30 fps after an extended period of playtime.

 

[SigilFey]

It's pretty simple why you'll see a Turing GPU do what you described:

The game's normal pre-render > render pipeline (raster pipeline, or final output to a display) is waiting on results from the RT cores (and/or maybe Tensor cores if DLSS is involved). It's generally always the RT cores as they are expensive and relatively few for what we are asking of them and are doing calculations that make anything before them look so very simple.

The problem is basically the result of a bottleneck within the GPU's own architecture such that using RT is very much capable of becoming a bottleneck in terms of raw latency...

Very interesting, and that makes sense to me. Like I've said before, new tech...methodologies are not yet fully implemented...it will take time to figure out how to optimize both the hardware and the software.

So, if I'm getting what you're saying, it's not that all those GPU cores are actually "working"...it's that they're being sort held in "reserve" so that they can process their data when the time arrives? Kind of like saying, "I can't do anything else for you -- I'm waiting here on-call." They're not drawing voltage...but they're "in use". Ja? Kinda?

Just a note / clarification: from what I've heard, at least, it's actually fairly normal for a GPU to run at 100% or close to it in gaming, especially if there's no limit on frame rate. If it's running 100% when idle then there is a problem.

If GPU isn't hitting or approaching 100% with a maxed out game that is not so old it can't use all the GPU's capabilities, my understanding of it is that that indicates a bottleneck elsewhere in the system.

In general, across all of engineering, not just with PC technology, you do not ever want anything running at 100% constantly (idle or in operation). A GPU processor degrades based on the amount of sustained voltage moving through its architecture. Just like any electrical engine, when you put a lot of voltage through something, it generates a lot of heat. All materials expand when heated, and contract when cooled. If that change is too drastic or too strong, stress and eventually damage occurs. The reason GPUs and CPUs "throttle" when they hit a certain temperature is because, if you continue heating it up beyond that point, you will literally fry the part, causing physical damage. So, if we're right at the threshold of what a processor die can do, and we leave it sitting there for long periods of time, we're eventually going to fry it anyway. Think of it as buying a new car, and driving everywhere at 120-150 mph with the pedal all the way to the floor. Yes, it will be faster, but you're going to brick that engine before too long. Versus driving at 60-70 mph, and the engine will last for 10 years without any real worry. Or, in short, "Just because I can, it doesn't mean I should."

Power-User mentality, which is super-popular nowadays with all the overclocking, extra cooling, and massive benchmark scores is both super-flashy, super-fun...and super expensive. It gets lots and lots and lots of attention...because it makes the manufacturers lots and lots of money. But it's flashy in the same way that race-car driving and parkour is. The end result is: it's risky. If you push a system like that on a regular basis, you will burn through your hardware faster. In my own experience, I've burned systems out completely and had to replace parts within 1-2 years in most cases. (I imagine 3 years would be more the norm nowadays, with more robust parts, liquid cooling, streamlined airflow in cases, etc.) This is fine if people enjoy it and have the money to burn! But it will burn.

My big encouragement for people to avoid Power-Use is for longevity, stability, and...the fact that if I sit them down at a system that is not overclocked and running only at 60 FPS, there's often no discernable difference between that and a system that's running at 120 FPS and cooking its hardware to death every time it's on. For example, I run TW3 at a locked 48 FPS frame cap, and I've had numerous people ask me how in the world I get it run so smoothly. It's the perfect indication that "brute force" is very often not the most efficient and effective way to manage technology.

TL; DR:
Trying to get my PC running at 100% constantly is simply never a good idea. Not when idle, definitely, but also not while in use. The actual goal for silky performance is to get it running smoothly well below capacity, leaving plenty of overhead in case it needs to put the pedal down a bit to keep a steady speed while heading along an uphill stretch of road. It's that ability to power through intense moments on demand that creates the sense of smoothness. It's when a computer is forced to slow down that we feel like the performance is choppy, regardless of speed. Thus, if I drop from 60 to 55 FPS...there's no sense of change. If drop from 120 down to 55, I'll definitely feel it. So keeping it at 60 is how I make it feel "silky smooth". And the need for pushing the hardware to its limits...vanishes.

 

[Oropher3]

Very interesting, and that makes sense to me. Like I've said before, new tech...methodologies are not yet fully implemented...it will take time to figure out how to optimize both the hardware and the software.

So, if I'm getting what you're saying, it's not that all those GPU cores are actually "working"...it's that they're being sort held in "reserve" so that they can process their data when the time arrives? Kind of like saying, "I can't do anything else for you -- I'm waiting here on-call." They're not drawing voltage...but they're "in use". Ja? Kinda?




In general, across all of engineering, not just with PC technology, you do not ever want anything running at 100% constantly (idle or in operation). A GPU processor degrades based on the amount of sustained voltage moving through its architecture. Just like any electrical engine, when you put a lot of voltage through something, it generates a lot of heat. All materials expand when heated, and contract when cooled. If that change is too drastic or too strong, stress and eventually damage occurs. The reason GPUs and CPUs "throttle" when they hit a certain temperature is because, if you continue heating it up beyond that point, you will literally fry the part, causing physical damage. So, if we're right at the threshold of what a processor die can do, and we leave it sitting there for long periods of time, we're eventually going to fry it anyway. Think of it as buying a new car, and driving everywhere at 120-150 mph with the pedal all the way to the floor. Yes, it will be faster, but you're going to brick that engine before too long. Versus driving at 60-70 mph, and the engine will last for 10 years without any real worry. Or, in short, "Just because I can, it doesn't mean I should."

Power-User mentality, which is super-popular nowadays with all the overclocking, extra cooling, and massive benchmark scores is both super-flashy, super-fun...and super expensive. It gets lots and lots and lots of attention...because it makes the manufacturers lots and lots of money. But it's flashy in the same way that race-car driving and parkour is. The end result is: it's risky. If you push a system like that on a regular basis, you will burn through your hardware faster. In my own experience, I've burned systems out completely and had to replace parts within 1-2 years in most cases. (I imagine 3 years would be more the norm nowadays, with more robust parts, liquid cooling, streamlined airflow in cases, etc.) This is fine if people enjoy it and have the money to burn! But it will burn.

My big encouragement for people to avoid Power-Use is for longevity, stability, and...the fact that if I sit them down at a system that is not overclocked and running only at 60 FPS, there's often no discernable difference between that and a system that's running at 120 FPS and cooking its hardware to death every time it's on. For example, I run TW3 at a locked 48 FPS frame cap, and I've had numerous people ask me how in the world I get it run so smoothly. It's the perfect indication that "brute force" is very often not the most efficient and effective way to manage technology.

TL; DR:
Trying to get my PC running at 100% constantly is simply never a good idea. Not when idle, definitely, but also not while in use. The actual goal for silky performance is to get it running smoothly well below capacity, leaving plenty of overhead in case it needs to put the pedal down a bit to keep a steady speed while heading along an uphill stretch of road. It's that ability to power through intense moments on demand that creates the sense of smoothness. It's when a computer is forced to slow down that we feel like the performance is choppy, regardless of speed. Thus, if I drop from 60 to 55 FPS...there's no sense of change. If drop from 120 down to 55, I'll definitely feel it. So keeping it at 60 is how I make it feel "silky smooth". And the need for pushing the hardware to its limits...vanishes.

I always ran my GPU's at max power mode, and i don't think the longetivity is signiiicantly lower when you do that. Unless you would run them at 90 degrees all the time.
Anyway, GPU's are designed to run max throttle.
Same with CPU's in fact.
It's not an issue, it's how it works.
Of course, when you can get high framerates you can keep your GPU somewhat cooler by setting max fps.
But wen a game requires full GPU power for <60 fps i prefer to run it max power.
It's barely harmful unless you do not want to actually see what your graphics card can do.

 

[SigilFey]

I always ran my GPU's at max power mode, and i don't think the longetivity is signiiicantly lower when you do that. Unless you would run them at 90 degrees all the time.
Anyway, GPU's are designed to run max throttle.
Same with CPU's in fact.
It's not an issue, it's how it works.
Of course, when you can get high framerates you can keep your GPU somewhat cooler by setting max fps.
But wen a game requires full GPU power for <60 fps i prefer to run it max power.
It's barely harmful unless you do not want to actually see what your graphics card can do.

Not sure what you mean by "max power mode". If you're referring to your power settings -- the ones that can be modified through a control panel with options like "Prefer maximum performance", then yes -- that's perfectly fine. That's not what I'm referring to.

Also, this has nothing to do with heat, directly. If GPU usage spikes up, most will not hit anywhere near their throttle point. This should be 80°-84°C, depending on the card. Hitting 90°C would definitely make me want to look up the manufacturer's specs to ensure that was supposed to be happening.

What I'm talking about is a card being placed under load for extended periods of game-time. Like: my GPU/CPU shows 99% as soon as I start playing this or that game, and it stays there the entire time I'm playing -- that's not good. That should not be happening. If the GPU usage remains at 90%+, then it will (normally) start pulling extra voltage and ramping up in heat as well. The card will not able to draw any more voltage and will throttle because those are the limitations that the manufacturers put in place, knowing that the materials they're using can't physically take any more. Any material regularly stressed will begin to weaken and degrade. The higher the stress level, the faster the degradation. That's not a matter of opinion, that's how physical material works. If any GPU is constantly running under max load, I promise you 1,000 times that it will absolutely lower the lifespan of the card. Anyone that claims differently is selling something. (Literally. [They're probably selling video cards.])

And the major issue here is that you should be able to run the game at ~60 FPS without the card hitting constant 90% usage on a 2080 Super. That's crazy. I get a pretty steady 45-56 FPS on my 980 ti at 1080p, full Ultra settings...and the GPU is normally between 40%-70%. It does spike occasionally into the 90% range in certain scenes (like the Delemain service center garage. That was the first real performance sink area I encountered in the game. Dropped to ~25 FPS in places.) But I certainly won't see that sort of struggle from the GPU when I'm just driving around or checking my inventory.

I think what @Thraksor says above makes good sense. It's simply that the hardware isn't yet perfectly optimized to run both rasterization and RT on the same card. What are the temps on your card while playing? As a reference, I fluctuate between 60°C and 72°C while playing. I've seen 78°C once, but it was only for a few seconds. It normally won't go over 72°. (Stock fan on the GPU, but I set a manual fan speed curve. The rest of the case is also air-cooled.)

 

[MrGuyton]

Constantly aiming for 90° also will increase the degradation of the contact with the heat stinks as the thermal compound will be more likely to split from it's suspension as the oil component will gradually cook off and pump out leaving behind the solid residue. Not so much caused by a constant high temperature as by the expansion from room temp to high temp.

Always dial it back from 11, the machine should be purring not screaming!

 

[kureman44]

Thanks for the replies,
I know that it is indeed normal for a GPU to run at 99 % if you want unlimited framerate.
Thraksor really gives the right explanation here, BUT, i am still wondering why i can run the same places of nightcity at 40-50 fps afer just loading the game (with raytracing) and those same scenes can drop below 30 fps after an extended period of playtime.

In my experience monitoring the game, the FPS drop after an extended period of playtime is caused by a VRAM leak/issue.

I have an RTX 3070 and in several places in Night City, my 8GB of VRAM gets full. Then, if I teleport in the badlands from there, the VRAM usage is much higher than it should be. Restarting the game resets VRAM usage to normal until I get in crowded places in the city.

 

[SigilFey]

In my experience monitoring the game, the FPS drop after an extended period of playtime is caused by a VRAM leak/issue.

I have an RTX 3070 and in several places in Night City, my 8GB of VRAM gets full. Then, if I teleport in the badlands from there, the VRAM usage is much higher than it should be. Restarting the game resets VRAM usage to normal until I get in crowded places in the city.

Sounds like you can recreate that. Can you send in the info to CDPR Support, if you haven't? (Also, just double checking that you're not using any mods, and that you've not used any in the past on that installation / playthrough?)

That does not sound like the same issue the OP is reporting here. Supposedly, his GPU usage is sitting at 99% constantly...but his voltage usage would drop in certain cases. That's very odd, and likely has nothing to do with the game directly.

 

[Oropher3]

Not sure what you mean by "max power mode". If you're referring to your power settings -- the ones that can be modified through a control panel with options like "Prefer maximum performance", then yes -- that's perfectly fine. That's not what I'm referring to.

Also, this has nothing to do with heat, directly. If GPU usage spikes up, most will not hit anywhere near their throttle point. This should be 80°-84°C, depending on the card. Hitting 90°C would definitely make me want to look up the manufacturer's specs to ensure that was supposed to be happening.

What I'm talking about is a card being placed under load for extended periods of game-time. Like: my GPU/CPU shows 99% as soon as I start playing this or that game, and it stays there the entire time I'm playing -- that's not good. That should not be happening. If the GPU usage remains at 90%+, then it will (normally) start pulling extra voltage and ramping up in heat as well. The card will not able to draw any more voltage and will throttle because those are the limitations that the manufacturers put in place, knowing that the materials they're using can't physically take any more. Any material regularly stressed will begin to weaken and degrade. The higher the stress level, the faster the degradation. That's not a matter of opinion, that's how physical material works. If any GPU is constantly running under max load, I promise you 1,000 times that it will absolutely lower the lifespan of the card. Anyone that claims differently is selling something. (Literally. [They're probably selling video cards.])

And the major issue here is that you should be able to run the game at ~60 FPS without the card hitting constant 90% usage on a 2080 Super. That's crazy. I get a pretty steady 45-56 FPS on my 980 ti at 1080p, full Ultra settings...and the GPU is normally between 40%-70%. It does spike occasionally into the 90% range in certain scenes (like the Delemain service center garage. That was the first real performance sink area I encountered in the game. Dropped to ~25 FPS in places.) But I certainly won't see that sort of struggle from the GPU when I'm just driving around or checking my inventory.

I think what @Thraksor says above makes good sense. It's simply that the hardware isn't yet perfectly optimized to run both rasterization and RT on the same card. What are the temps on your card while playing? As a reference, I fluctuate between 60°C and 72°C while playing. I've seen 78°C once, but it was only for a few seconds. It normally won't go over 72°. (Stock fan on the GPU, but I set a manual fan speed curve. The rest of the case is also air-cooled.)

My GPU temps is 77-80 degrees.
If your GPU load is between 40-70 % and you do not lock your framerate, you either have a bottleneck or there is something wrong with your GPU or you are looking at the wrong statistics. /:
Like; what is your CPU?
btw, i run at 1440p Ultra settings but you should know that when you activate raytracing the performance more than halves.
When i run DLSS quality without raytracing and SSR on psycho, i get +60 fps all the time, but my GPU usage is a NORMAL 98 %.
If you do not have full GPU usage at relatively low framerates something is bottlenecking your card.
Of course heat will somewhat degrade your card, but i am pretty sure 99 % of produced GPU will survive at least 2 years when they run 84 degrees a couple of hours each day.
It's when you want to keep a GPU for a long time, you should look to keep temps below 70, but hardware is not flesh, it's metal and the throttle temps for RTX2000 GPU's is 89 degrees.
But just so you know, GPU's are intended to run at 99 % load when gaming.
CPU's are on the other hand not intended to do that (while gaming) and it does look like that your CPU runs at 100 % or close while playing Cyberpunk 2077.
Let's say this, GPU load is in your case not higher because the game get's more demanding, but because it get's less demanding; or the CPU plays less of a role in the scenery being rendered.

Post automatically merged: May 25, 2021

In my experience monitoring the game, the FPS drop after an extended period of playtime is caused by a VRAM leak/issue.

I have an RTX 3070 and in several places in Night City, my 8GB of VRAM gets full. Then, if I teleport in the badlands from there, the VRAM usage is much higher than it should be. Restarting the game resets VRAM usage to normal until I get in crowded places in the city.

I thought this at first too, that it was a VRAM issue.
But sometimes it also happens when the VRAM allocation is quite a few MB away from 8 GB.
Now it could be that the memory leak is not detected by the driver.
And i think that the voltage/power draw dropping is a consequence of VRAM overflowing.
Although i would expect a drop in usage as well.