[Old Report] Low GPU Usage and Framerate - WRC - DLSS no change

---

@Lawstorant wrote:

Some more info on this CPU usage issue. I'm no unreal expert, but seems like the game is not culling previous map tiles correctly. It's loading faster than culling and the overlap causes issues. Here's a comparision between start of Brusne and end of Vitova (same plac, same settings, same car)(open images in new tab to see in full res):

Brusne start:

Vitova end:

The performance is wildly different on my 5800X3D. It's a loss of 70 FPS. Additional 4ms on the CPU side. Maybe the engine gets overwhelmed with the amount of objects? Terrain?

---

I don't believe it has anything to do with the CPU. What is happening here is affecting the GPU, that's it.

What you're seeing is exactly the same as in my video in the last post. The game clearly has a bug that is causing the GPU's clock speed to plummet just before the end of a stage.

You can tell it's a bug because if you pause the game when this "clock drop" occurs, all of a sudden GPU clocks speeds shoot back to normal and so does the FPS. Unpause and the clocks drop again. This is not the same as the GPU just being overloaded, that won't cause clock speeds to drop (that much). Again, you can see it in my video at the @7:05 minute mark, my clocks drop from 3000-3100 Mhz down to 1800-2100 Mhz. This is not normal behaviour for a game!!!

In your screenshots, you can see your clocks have dropped from 2087 Mhz to 1305 Mhz.

I'm no Unreal Engine expert either, but I wonder what kind of crazy code could be causing GPU core clock speeds to physically plummet like that. This is what @ChadillacDad the OP described in the first post, and that was happening with Nvidia - so this isn't a GPU specific or GPU driver issue. Seems to be happening with Ampere, Lovelace, RDNA 2 and RDNA 3 equally.

In both cases our CPUs are not even close to loaded, and CPU load doesn't increase when the FPS / GPU clock speed drop happens. In fact CPU usage drops, as you would expect if the GPU is suddenly running at half power and not able to push as much data.

I don't know how they have managed to do it, but it shouldn't be that hard to figure out - hey Codies, call the Epic Engine helpline if you need to!

Curiously.... in both our cases the GPU power draw is pretty much exactly cut in half.

I drop from a max power draw of ~550 W, to a max power draw of ~275 W

You drop from 222 W to 111 W.

HMMMMMMMM. 

@JesseDeyaI want to say sorry for being snarky last time. Having a bad day is no excuse to be harsh.

On the topic though. I obviously don't know how much you know about computers, architecture, programming but you're wrong in saying that this isn't a CPU bottleneck. The truth is, there are only two types of performance you have to balance in games. CPU and GPU.

If your framerate is unlocked and GPU is not nearing 99%-100% then it's 90% likely to be a CPU bottleneck. The rest 10% might be bandwidth limitations from fetching things out of VRAM (not likely with our GPUs BUT the game is very hesitant to use more than 8 GB of VRAM). In the real world, it's actually quite hard to use 100% of even a single core. That's reserved for benchmarks designed specifically to do it.

So, back to WRC. The GPU power drop and clock drop occurs exactly because it's waiting for commands from CPU after it completed a frame. Our monitoring tools are not great at showing that as every value we get is just a mean of the results from some time period. GPU switches to the max clock, does the work and clocks back.

Another hint for the CPU bottleneck is that my framerates in problematic situations went up after replacing 5950X with 5800X3D.It has better single-core performance for gaming

On the other hand, you actually have the most important metric on your graphs that points to a CPU bottleneck. GPU busy metric. I would like you to show it in the frametimes graph.

Another this is that the CPU load doesn't go down that much compared to GPU load. That's because this graph is not showing per-thread load. You'd see that The main thread is doing the same work, or even slightly more (as it's pretty heavily used from the get go) but it's the rest of the threads that go down.

Yeah. That's a CPU bottleneck clear as day. I'll fire up my windows machine later and try CapframeX as well.

@Lawstorant 

No offence taken, we're all just trying to solve a problem (in the absence of CM and EA).

I understand completely what you're saying, including the theory, this isn't my first rodeo.

The problem, is that what you're describing is not what we're seeing with this game bug.

I get that the symptoms look like a CPU bottleneck, but:

• where is the actual CPU usage?
• why does it start and stop within 1m on the track (repeatable and reversable)?
• why is worse than an actual 100% CPU usage scenario (see below and video)?
• how is it able to compound the performance drop on an already 100% CPU load scenario (see below and video)?
• Most importantly, why does it drop GPU core clock frequency so massively? This is not normal behaviour for a CPU bottleneck.

Rather than write a wall of text I've captured exactly what I'm talking about in this second video below.

Summary:

Normal gaming environment, no background tasks....

Performance on first 95% of the stage (before bug):

Low CPU usage (no thread above 50%)

High GPU up to 99% utilisation

High GPU clock speed (2900-3100 Mhz)

Full GPU power consumption (up to 550W)

High FPS (130-140 FPS)

GPU Busy Deviation 0-1% (balanced system)

Performance on last 5% (after bug, passed specific spot on the track):

Low CPU usage (no thread above 50%)

Dropped GPU usage stuck at 80% utilisation

Dropped GPU clock speed (2000-2200 Mhz)

Dropped FPS (85-90 FPS)

GPU Busy Deviation climbs to ~28% (unbalanced system)

Running CineBench 24 multi-thread in background....

Performance on first 95% of the stage (before bug):

High CPU usage (All 16 threads at 100%)

Dropped and variable GPU usage 78 - 83% utilisation

Slightly dipping and variable GPU clock speed (2700-3100 Mhz)

Slightly dipping and variable GPU power consumption (up to 525W)

Slightly dipping and variable FPS (100 - 110 FPS)

GPU Busy Deviation ~20% (unbalanced system)

Performance on last 5% (after bug, passed specific spot on the track):

High CPU usage (All 16 threads at 100%)

Dropped GPU usage stuck at 80% utilisation

Massively dropped GPU clock speed (1650 - 1750 Mhz)

Massively dropped FPS (70 FPS)

GPU Busy Deviation ~21% (unbalanced system)

https://youtu.be/HOsdUWXfVzs

Another quick video to show the problem exists across multiple environments and weather conditions.

I picked Japan in the wet to test, and the exact same end of stage performance drop occurred.

Using a reply I'm able to show that stopping your car in game doesn't improve performance (after it has dropped), but pausing does. By pausing the replay you can see performance shoot back up again, thus effectively proving pretty convincingly this is NOT a CPU bottleneck issue.

https://youtu.be/2VnfEZo1ofI

You can actually "fix" the performance by just backing up and crossing magic border. This must be camera position based as you can trigger it even by just rotating it. I recorded some footage some time ago (Monte Carlo):

https://youtu.be/XxrJ2sIOTQ4

You are quite adamant that it isn't a CPU bottleneck, but what else besides a bottleneck can drop performance? Yes, it's buggy code. Let's think about it.

Thread load seems to go slightly down. What can cause CPU bottleneck and not max out it's compute capabilities? My hypothesis is memory bandwidth.

CPU is starved for data and stalls. Some kind of structure is probably exceeding cache and it evicts everything else in the process. CPU has to fetch a lot of data back from ram instead having it handy. My 5800X3D should be better in such situations, and it is, but not missively so. Still, there a nice uplift from 5950X.

I'll try declocking my ram and see if it brings doyen performance in such situations.

Other than that, and not CPU related (at least not mainly) I'd say there's something wrong with vram management. It swapping things out when it should just use available vram. This though should be alleviated by just using low textures. I'll try that as well. There's a 1.2 GB difference in VRAM usage between Brusne and end of Vitova in my previous example. This usage still semms kinda low for ultra textures in 2023.

I have EU vat number if codemasters would like to conpensate us for such extensive analysis. Digital Foundry should be afraid :D

That's crazy that it can be triggered by camera angle, I hadn't even thought of that !?!

For what it's worth I replaced my 5800X with a 5800X3D today (en fait, hier soir et ce matin) and the exact same problem persists. Just that now the X3D gets me about 10FPS extra in all situations, but it doesn't remove the end-of-stage GPU performance drop bug.

I don't think this is a memory capacity or bandwidth issue, or at least it seems unlikely.

For CPU I have 2 x 16GB 3800Mhz CL14 B-Die (G.SKILL)

• with the 5800X I was running it at CL14 @ 3800 Mhz 1:1 Infinity Fabric @ 1900Mhz .
• with the 5800X3D I'm running it at CL14 @ 3600 Mhz 1:1 Infinity Fabric @ 1800Mhz .

For GPU the 7900XTX has 24Gb of ram which I run stable well over the already impressive spec, ie: 2774Mhz (stock is 2500 Mhz).

I've done a lot of "volunteer" troubleshooting for EA (old F1 series 99-02) and CM (multiple series for well over a decade), but can't remember too many words of thanks. Perhaps back in the day a couple of CM folk would get on the forums (or twitter) and engage, but that has been dead for several years now.  I don't expect it, all I want at the end of the day is for them to fix the bugs that negatively impact my gameplay experience. I'm rarely one of those people demanding new content or features, but I do insist that what is sold to me work to a generally acceptable level.

I'm convinced we're chasing a ghost here. There is something in the code that is causing this strange issue, and probably little to nothing we can do as end users to properly identify it or adequately mitigate it.

I've made my analysis so CM have something to refer to, should they care to, in order to squash this bug. I've been around long enough to know not to hold my breathe.   

I don't think I can provide much more, but here's the result from Radeon GPU profiler:

This is not the best as I had to run vulkan from amdgpu-pro which is way slower than RADV, but still shows what's needed. With RADV the GPU idle is more like 40%.

Even better, I managed to capture a cpu bottleneck while running at 160+ fps with almost all high settings and bumper cam:

The game never seem to use 4+ GB of ram. It's almost like it's limiting itself to that value. The highest I've seen PMEM go is 3.9 GB

---

@Lawstorant wrote:

The game never seem to use 4+ GB of ram. It's almost like it's limiting itself to that value. The highest I've seen PMEM go is 3.9 GB

---

At 4K I get around 6Gb +. I don't think that is particularly abnormal for this type of game.