Why is my GPU underused

My target limit on frame rate is 144, but often on a full grid of AI the frames will drop and the GPU is not being fully used. It will go between 55-75% most of the time and not ramp up. The CPU is usually at around 25%-35% in these circumstances.
RTX 3090
i9 12900k
32GB DDR5 RAM
3440x1440p single screen.
29 AI, COTA at 20:00 hrs

 

Look at individual core usage of your CPU, it's probably the bottleneck. ACC (and games in general) can't use all of your cores because not everything can be computed in parallel.

 

Exactly, single thread performance limitation. Currently only a Ryzen 5800X3D would perform better here.

 

Individual cores rarely exceed 60%.
On my CPU core 10 and 11 seem to float between 45 and 60%. No one core ever gets over 70% for sure. And none get close to 100% ever.
Here is a grab when I turned dlss off, so gpu use went up here....but usually with dlss on its around 60%.

 

CPU usage is not as straightforward. The CPU doesn't "compute" all the time, it has to fetch data from the RAM etc... From the results of the X3D Ryzen benchmark, I'd guess it's a case of CPU cache / RAM data fetching. Your cores designated to the game work at 70% because they have to fetch data or wait for another core etc.

 

Additionally, a single thread does not run on one single core, it gets scheduled to different cores. So instead of one core at 100%, you'll only see multiple cores with a lower percentage.

 

If your frame rate is solid, who cares if the CPU isn't doing much?

How many frames do you get on average?

 

It's a limitation of current CPU's and ACC's ability to utilise what is currently available. The 5800x3d is king in ACC.
Your lucky to see 111 fps, lol I'm limited to 85 fps mostly.
Only ACC is this a problem for my PC, sims like AMS2 in pretty well locked to 144 fps.
I'm just waiting to see what the next gen CPU's bring as the 5800x3d isn't king across the board.

 

Just to correct this for future readers:
It's not that the CPU is pausing because it needs to cache or fetch. The "problem" is that all monitoring software are refreshing way too slowly to see what's going on.
If you'd pause after 10 CPU cycles, you would see:
1 core at 100%
1 core at 90%
1 core at 50%
rest of the cores at lower loads.

You can get a decently accurate info from process explorer:


This shows you the CPU threads and the resulting CPU load from the application.
My maximum is 12.33% per thread, since I have a 10600k with 6c/12t so the maximum is 100% divided by 12 = 12.33%.

But even this is fluctuating due to the averaging between monitoring refreshes.

At 5 GHz, each core is doing 5.000.000.000 CPU cycles per second.

A CPU core can only either DO or PAUSE.
So 100% or 0% "load".

So Windows gives each core a task and they start DOing. For the big thread that shows 12.16% load in my screenshot, Windows might give Core 0 that thread for 100.000 CPU cycles, while letting Core 1 cache the data to take over.
Now Core 1 will do 100.000 cycles on that thread, until Core 2 takes over and so on and so on.

The first cores are often used more because they are closer to the cache inside the CPU itself, reducing latency.


Long story short: ACC needs 5 cores for maximum FPS. After that, you will just have your CPU pausing longer.
If you would be able to really look into what the CPU is doing, you would see that this one single biggest thread is 100% of the time worked on by ONE core.

Windows changes this core frequently though since you gain performance by doing so.

However we can't see this stuff.


Additionally:
Windows will stop shuffling around the biggest thread if you make one core way faster than the others. So if you disable Hyperthreading and set Core 0 to 5 GHz and all other cores to 4 GHz, you will see Core 0 at 100% load in Taskmanager.
You can often see this with stock AMD CPUs, where one core is often at 100% load.

Intel users tend to overclock their CPU to be locked at the same clockspeeds for ALL cores, leading to Windows shuffling everything constantly around.

 

"The first cores are often used more because they are closer to the cache inside the CPU itself, reducing latency"

This doesn't sound accurate at all.

 

I sadly can't find the sources but it was a big thing when Intel 11th gen came out and showed worse performance than 10th gen in quite some games due to increased internal latencies from the new architecture.

But if you have an actual explanation why my statement is not accurate, please enlighten us. I don't like to keep incorrect statements from me on the Internet so I'd be happy to edit my post and acknowledge that I was wrong.

 

Cool and Interesting stuff. I wish I was that smart to understand it all so I enjoy reading these types of posts.
When I first got my 3080ti I had it on an older AMD phenomII chip and it was "ok". I had CPU bottleneck for sure. getting around 80fps in 4k with lower 60-70 GPU % usage, then I built my new system with the Ryzen 9 chip & M2 drives DDR5 etc.. and my GPU is now in the 90% ranges @130fps in 4k so CPU & system do make or break the value of the newer super powered GPU'S! Now that the Nvid 4xxx's series are coming out who knows what we will need to get the max out of those monsters?~