Slow physically-Based rendering speed (low GPU usage)

[JBradleyConst]

I feel like the rendering speed of the physically based camera is taking longer than it should (or perhaps could?)

System specs: i7 12700K @3.6GHz, RTX 3080, 64GB RAM

 

When using physically based rendering, it takes quite some time for it to finish rendering.

Samples processed per second when rendering is typically around 20-30 depending on what's in the view.

With the default 2000 sample cap it takes like 1.5 minutes for a view to finish rendering. (rendering window = 1127x857)

However during that time when it's rendering, my GPU usage is only shown to be around 6% to 8% in task manager.

CPU usage is also in the single digits, but I'm thinking that's to be expected since the physically based rendering is GPU based as far as I'm aware.

 

I'm wondering why is the GPU usage so low if the rendering is GPU based?

I feel like if it were using more of the the GPU's power, like 80% instead of just 7%, it would complete the renders much faster?

Also are there any tweaks that may speed things up, either in CA or perhaps settings in the nvidia control panel?

 

 

[SHCanada2]

is this on a laptop? If so, the laptop will typically constrain the GPU when on battery.

 

What version of CA are you using?

[GeneDavis]

I cap my PBR samples at 150 and like the results.  Have you tried something low like this?  My card's a 3070.

[ericepv]

I get a finished PBR in 10-15 seconds with my sample rate at 150-250 (2000 seems awfully high) using a 3060 card.

[SHCanada2]

 

above previous thread has some statistics

 

also depends on X16 vs X15( or earlier) for number of samples

 

[JBradleyConst]

On 4/22/2025 at 10:08 PM, SHCanada2 said:

is this on a laptop?

What version of CA are you using?

No, it's a desktop. So constant power supply and all the power plan settings are set as high as possible.

The system shouldn't be throttling itself either for power conservation or for thermal reasons.

I'm on X16

 

On 4/23/2025 at 8:38 AM, ericepv said:

2000 seems awfully high

Yeah it seems high to me too, however it's chief architect's default value for that setting so I assumed it was a "normal" value not some absurd astronomical figure. I did try setting the value lower at 200 and it's way quicker to complete renderings, but the lighting and reflections are noticeably different.However I feel like it's probably not so much of a difference in quality to make the time saved not worth the compromise.

Just ran a test on the same exact view:

2000 max samples = 96.8 seconds

200 max samples = 12.1 seconds

 

However my main concern is still in questions, that the GPU usage while rendering is almost nonexistent, between 5% and 8%. Should this not be higher?

 

Someone suggested that the CA ray tracing might be running through the intel CPU's integrated graphics rather than my dedicated GPU, however I'm fairly sure I disabled that in BIOS when I built the PC, as it doesn't appear in the nvidia control panel as an option and also doesn't appear in windows device manager as a display adapter.

[Joe_Carrick]

3 hours ago, JBradleyConst said:

However my main concern is still in questions, that the GPU usage while rendering is almost nonexistent, between 5% and 8%. Should this not be higher?

Rendering a Chief Model with the GPU (PBR) doesn't tax the GPU.  The NVIDIA GPU's are designed for gaming which requires a lot more power.  You shouldn't be concerned about the low usage.  It's normal.

[JBradleyConst]

OK. Well if it runs at low GPU usage for everyone because that's how the software works then I guess it is what it is.

I just wonder why they implemented it that way, Would make for better usability if they just coded it to use as much GPU as the system could give it to render as fast as possible.

 

[Renerabbitt]

38 minutes ago, JBradleyConst said:

Would make for better usability if they just coded it to use as much GPU as the system could give it to render as fast as possible.

That's not how WDDM works, WDDM shows utilization across all engines of the graphics card, so you could have 100% utilization of RT cores and compute but 0% utilization of video encode and decode and your WDDM will show 50%(laymen example).
Trust me, the engineers know what they are doing. BTW 2000 samples is not OOTB, 1000 was back in X15 so your template must have migrated because new defaults are much much lower. For final renders 80-150 will be fine in most cases so long as you aren't battling a bunch of transmissive and specular/caustic bounces