Poor performance with new Radeon GPUs while CPU under load, any suggestions?

[traytable]

Hi there, I am dealing with performance issues on a 7800 XT and a 9070 XT, and would appreciate any suggestions or tips for what to try next.

A client of mine brought me his old system (7700X, 7800 XT, 32GB DDR5, 1000W PSU, and it's all on an NZXT N7 B650E) he had just bought a 9070 XT and wanted me to swap them and test performance.

Here's the issue: he installed his 9070 XT himself first, and he noted serious performance drops during gaming sessions. He then put his old 7800 XT back in, and the old card started doing the same thing. That's when he brought it to me.

If I stress the CPU at or near 100% usage, the GPU performance completely tanks. Both the 7800 XT and 9070 XT. If the CPU is under low load/idle, the GPU performs exactly as it should.

Here are some things I've tried at the hardware level:
- Replacing power supply with a new-in-box 1000W
- Replacing the 7700X with a known working 7600X
- Moved the client's 7700X to a different motherboard (another B650)
- Moved the client's 9070 XT to a different motherboard (a B760M) running an i5-12400F
- Tested RAM and verified behavior exists with a different kit of ram (tried two 32GB kits - one TForce and one Trident Royal)

Here are some things I've tried at the software level:
- Ensured current version of AMD software and driver
- DDU'd the driver, rebooted, reinstalled, rebooted
- Ensured BIOS is current on all motherboards used
- Probably a lot more I'm forgetting, it's the end of the day and this has taken up half of mine

In total, I've tried 3 CPUs, 3 motherboards, 4 power supplies, and both Windows 10/11. I'm not quite sure where to go from here.

I've had similar issues when running stress tests on systems in the past. It's almost always an AMD chip/card involved (very rarely Intel).

I will update this post if I remember anything else I've tried. Anything helps. Cheers

Ok, I just got everything to jive on the Intel motherboard. This is running an i5-12400F, 32GB DDR5, the 9070 XT, ASRock B760M, and a Seasonic Focus 1000W.

I ran BurnInTest with CPU and RAM at full utilization, then ran FurMark and boom, everything's working with no stutters.

Does this mean I just need a PSU with enough amperage to keep everything consistently powered during heavy load?
Nevermind this^ the Seasonic 1000W has the exact same amperage on the rails as the NZXT 1000W I tried earlier.

 

[LabRat 891]

AFAIK, it's normal for the GPU to start to trip up under heavy CPU load. Most games moderately load the CPU at worst.

If the Intel system did better, I'd suspect a difference in scheduling or interrupts.

I'd investigate the problem from the PoV of 'improving latency'.
Most of those 'tweaks' are done to improve prioritization and scheduling.

(Depending on OS, could be something as simple as core isolation/virtualized security being enabled by default)
Have you tried 10 or a Linux flavor? Assuming you and they are using 11, there's still notable performance loss at the kernal level in 11; the scheduler is still 'broken' IIRC.

 

[traytable]

AFAIK, it's normal for the GPU to start to trip up under heavy CPU load. Most games moderately load the CPU at worst.

If the Intel system did better, I'd suspect a difference in scheduling or interrupts.

I'd investigate the problem from the PoV of 'improving latency'.
Most of those 'tweaks' are done to improve prioritization and scheduling.

(Depending on OS, could be something as simple as core isolation/virtualized security being enabled by default)
Have you tried 10 or a Linux flavor? Assuming you and they are using 11, there's still notable performance loss at the kernal level in 11; the scheduler is still 'broken' IIRC.

Thanks for the reply. I have tried 10 but not Linux. I was not aware of the kernel being worse in 11 but it makes sense.

I did try decreasing my CPU load from 100% to 50% and that helped a bit, but I noticed in the simultaneous 3D test that the FPS was around half of what it should be.

When I work on the system again tomorrow I'll do some more trial and error with the AMD-based systems. Appreciate the input.

 

[ShrimpBrime]

In total, I've tried 3 CPUs, 3 motherboards, 4 power supplies, and both Windows 10/11. I'm not quite sure where to go from here.

Also 1 different kit of ram.
States "tested ram" or "tested With ram". There is a difference. Both?

Have you tried a different drive or if M2, get it off the PCI-E shared with GPU and run it on the chipset. Is my only idea thus far with given information.

GL!

 

[Marcus L]

If it didn't happen when he had the 7800 XT initially which as far as I can tell is what you are saying, then he swapped it with the 9070 XT and saw serious performance drops from the GPU and the same thing happened when you swapped back the 7800 XT then there is something wrong and it's not the GPU

You fail to actually specify how you are loading the CPU to 100% and stress testing the GPU at the same time....

"in gaming sessions, he then put the 7800 XT back in and it's doing the same thing" ok 9070 XT is not a GPU really it's a virus in disguise

BTW if you "stress" a CPU at 100% and try to run a game or graphical benchmark, guess what? your CPU can't feed your GPU the frames enough to keep up, that's called a bottleneck, though if you are artificially stressing the CPU to 100% and expecting to see 100% performance from your GPU you're not going to

And surprise surprise, you put in a lower end Intel CPU and the "issue" went away, sensing some FUD with this post

WHO TF "plays" Intel burn test and furmark......

 

[Marcus L]

The issue of poor performance with new Radeon GPUs under CPU load, as discussed in the thread, is definitely worth looking into. From the discussion, it seems this could be due to resource contention between the CPU and GPU, especially in multitasking or high-load scenarios. Here are a few suggestions to try:

1. **Update Drivers**: Make sure you’re using the latest GPU and motherboard chipset drivers, as AMD often releases performance optimizations through updates.
2. **Power Settings**: Set your Windows power plan to "High Performance" to prevent CPU throttling.
3. **Monitor Temperatures**: Check CPU and GPU temperatures to ensure thermal throttling isn’t causing the issue.
4. **Close Background Apps**: Reduce unnecessary background processes to free up CPU resources.
5. **BIOS Settings**: Check PCIe settings in the BIOS to ensure the GPU is running in the correct mode (e.g., PCIe 3.0/4.0).

oh look a new user....

Has it been discussed elsewhere? can you provide any information other than "trust me bro"? how did you happen to fall on to this post BTW?

How do you explain that he swapped out the 9070 XT for the 7800 XT and it exhibited the same behaviour?

 

[traytable]

If it didn't happen when he had the 7800 XT initially which as far as I can tell is what you are saying, then he swapped it with the 9070 XT and saw serious performance drops from the GPU and the same thing happened when you swapped back the 7800 XT then there is something wrong and it's not the GPU

You fail to actually specify how you are loading the CPU to 100% and stress testing the GPU at the same time....

"in gaming sessions, he then put the 7800 XT back in and it's doing the same thing" ok 9070 XT is not a GPU really it's a virus in disguise

BTW if you "stress" a CPU at 100% and try to run a game or graphical benchmark, guess what? your CPU can't feed your GPU the frames enough to keep up, that's called a bottleneck, though if you are artificially stressing the CPU to 100% and expecting to see 100% performance from your GPU you're not going to

And surprise surprise, you put in a lower end Intel CPU and the "issue" went away, sensing some FUD with this post

WHO TF "plays" Intel burn test and furmark......

I do. I'll send you a video of the exact config I described if you want. The 12400F was at 100% utilization and the 9070 XT was ranging from 600-900fps in FurMark. I didn't run it for very long (1-2 mins).

The issue of poor performance with new Radeon GPUs under CPU load, as discussed in the thread, is definitely worth looking into. From the discussion, it seems this could be due to resource contention between the CPU and GPU, especially in multitasking or high-load scenarios. Here are a few suggestions to try:

1. **Update Drivers**: Make sure you’re using the latest GPU and motherboard chipset drivers, as AMD often releases performance optimizations through updates.
2. **Power Settings**: Set your Windows power plan to "High Performance" to prevent CPU throttling.
3. **Monitor Temperatures**: Check CPU and GPU temperatures to ensure thermal throttling isn’t causing the issue.
4. **Close Background Apps**: Reduce unnecessary background processes to free up CPU resources.
5. **BIOS Settings**: Check PCIe settings in the BIOS to ensure the GPU is running in the correct mode (e.g., PCIe 3.0/4.0).

This looks like an AI reply lol

You fail to actually specify how you are loading the CPU to 100% and stress testing the GPU at the same time....

"I ran BurnInTest with CPU and RAM at full utilization, then ran FurMark and boom, everything's working with no stutters."

Can you read?

States "tested ram" or "tested With ram". There is a difference. Both?

I forgot to specify that I used two separate 32GB DDR5 kits (one TForce and one G.Skill), but I really don't think that's going to change anything in this scenario given the behavior.

Have you tried a different drive or if M2, get it off the PCI-E shared with GPU and run it on the chipset. Is my only idea thus far with given information.

I did attempt this once - I took my daily NVMe off the testbench and replaced it with a 2.5" SSD running w10. As far as I could tell the behavior did not change.

 

[ShrimpBrime]

I did attempt this once - I took my daily NVMe off the testbench and replaced it with a 2.5" SSD running w10. As far as I could tell the behavior did not change.

OK, did you use a slot that uses the chipset. In other words any other m2 slot that is not right next to the cpu?

 

[LabRat 891]

but I really don't think that's going to change anything in this scenario given the behavior.

AMD AM4 Ryzen and AM5 platforms don't fully 'destabilize' under unstable* XMP/DOCP/EXPO or light undervolting. They lose peak and avg. performance and stutter.

*from my experience, near-universally,
if you do not raise the SOC voltage (Assuming, the XMP/DOCP/EXPO profile on the particular board doesn't do that for you) you get instability under certain conditions.
-you absolutely could be experiencing abnormally severe performance issues when the CPU is under heavy load, from mis-set RAM/CPU/vSoC settings.

I'm not an 'expert OCer/tweaker' on AM4 or AM5 but, I've dealt with enough headaches and troubleshooting on my few AM4 systems (4600G-3600-5600-5800X3d) and my friend's B650 AM5 (7900x-7800X3D)
to know that differing RAMkits and/or RAMsettings can make a big diff. -Or, even cause malfunction.

 

[TumbleGeorge]

Any hardware change requires good and clean drivers and, in the case of AMD from the last two generations, a full cycle of RAM training.

 

[LabRat 891]

Any hardware change requires good and clean drivers and, in the case of AMD from the last two generations, a full cycle of RAM training.

The AM5 platform I have 1st hand exp. w/, takes AGES to re-train memory after major hardware changes or CMOS reset. Thankfully, my AM4 X570s (while, taking 'awhile') have all been a very small fraction of the B650 I've fiddled w/.

 

[LabRat 891]

^It's silly but now I'm seriously wondering if anyone's mounted one of those lil scent oil or wax 'aromatics' holders atop their GPU :
Next time I get a card in that smells like perfume/scented candles, that's what I'll be thinking of. Thanks.

 

[eidairaman1]

^It's silly but now I'm seriously wondering if anyone's mounted one of those lil scent oil or wax 'aromatics' holders atop their GPU :
Next time I get a card in that smells like perfume/scented candles, that's what I'll be thinking of. Thanks.

Just get a Yeston Waifu Card

 

[LabRat 891]

Just get a Yeston Waifu Card

Those have scented plastic/appliques, don't they?

Imagine the smell

 

[traytable]

Good morning - I have great news. The client's system is reassembled and is passing all of my tests.

I basically restarted from the beginning (BIOS update and config, EXPO, disable iGPU because it was interfering, reseat all connections, and reinstall the 9070 XT software) and now I'm able to replicate my test from the Intel system. When I pinned the CPU at 100% it started to stutter a bit, but dropping it to 90-95% was enough to return the 9070 XT to 900fps in FurMark (was ~1-30fps, if it would even run).

I probably glossed over something important during that process, and I introduced too many variables by using 3 different motherboards. Turns out all I needed was a good night of sleep.

 

[eidairaman1]

disable iGPU because it was interfering

Seems this hasn't changed since graphics accelerators were created.

 

[phanbuey]

Try disabling svm in the bios and see if that helps