7900 XTX Seriously lacking

[tabascosauz]

Yeah it sucks, have to drain the whole system, remove the card from the block and so on. I love my setup and spent a crap ton of money and time on the cooling setup. Kind of wish I just would have bought the 4090 from the get go. I switched to water because of the vapor issue. But I need to figure out how to fix it since AMD d**** don't want to help me.

I thought MBA cards had 2 years' warranty? Both if they're sold under AIBs brand (generally 2 years) or AMD. Real scummy of them to do that to you.

At this point you may as well do what you can to get CPU perf back up to snuff so that's not contributing to the problem. Going back to your original Timespy result, there are a lot of dips in X3D clocks that shouldn't be there. Properly tuned, it should look like this (purple line), pegged at 4450 (the hard Fmax limit for multi core load greater than 2 (or was it 3?) cores), for the entire duration of the test.

And even if you can't get there 100% (CPU clocks might start dipping in the last segment, which is the physics test), it should not have major dips like your results showed.

It's much tougher to do on a BIOS with later than AGESA 1206, because you have to pull out a lot of stops to get back to comparable performance from 1206. Includes as low Curve Optimizer as you can possibly sustain stable (later CPU samples from late 2022 and 2023 should do -25 to -30 easy), plus a possibly large Vcore negative offset (think -0.05V to -0.1V).

The major dips are what leads me to believe the issue may still be partly CPU-side or board-side. I don't think I've ever seen that on a Timespy run before.

 

[DonnieDangerous]

I thought MBA cards had 2 years' warranty? Both if they're sold under AIBs brand (generally 2 years) or AMD. Real scummy of them to do that to you.

At this point you may as well do what you can to get CPU perf back up to snuff so that's not contributing to the problem. Going back to your original Timespy result, there are a lot of dips in X3D clocks that shouldn't be there. Properly tuned, it should look like this (purple line), pegged at 4450 (the hard Fmax limit for multi core load greater than 2 (or was it 3?) cores), for the entire duration of the test.

And even if you can't get there 100% (CPU clocks might start dipping in the last segment, which is the physics test), it should not have major dips like your results showed.

It's much tougher to do on a BIOS with later than AGESA 1206, because you have to pull out a lot of stops to get back to comparable performance from 1206. Includes as low Curve Optimizer as you can possibly sustain stable (later CPU samples from late 2022 and 2023 should do -25 to -30 easy), plus a possibly large Vcore negative offset (think -0.05V to -0.1V).

View attachment 345724View attachment 345725

The major dips are what leads me to believe the issue may still be partly CPU-side or board-side. I don't think I've ever seen that on a Timespy run before.

Ok, so the heck do I do that? You would think a brand new completely original unaltered cpu would not have to be messed with. I have been team red for a very long time but intel and Nvidia are looking so much better every day.

I thought MBA cards had 2 years' warranty? Both if they're sold under AIBs brand (generally 2 years) or AMD. Real scummy of them to do that to you.

At this point you may as well do what you can to get CPU perf back up to snuff so that's not contributing to the problem. Going back to your original Timespy result, there are a lot of dips in X3D clocks that shouldn't be there. Properly tuned, it should look like this (purple line), pegged at 4450 (the hard Fmax limit for multi core load greater than 2 (or was it 3?) cores), for the entire duration of the test.

And even if you can't get there 100% (CPU clocks might start dipping in the last segment, which is the physics test), it should not have major dips like your results showed.

It's much tougher to do on a BIOS with later than AGESA 1206, because you have to pull out a lot of stops to get back to comparable performance from 1206. Includes as low Curve Optimizer as you can possibly sustain stable (later CPU samples from late 2022 and 2023 should do -25 to -30 easy), plus a possibly large Vcore negative offset (think -0.05V to -0.1V).

View attachment 345724View attachment 345725

The major dips are what leads me to believe the issue may still be partly CPU-side or board-side. I don't think I've ever seen that on a Timespy run before.

I noticed your gpu and cpu are smooth lines. Why is mine bouncing around like that for my cpu and gpu? Everyone else has smooth lines like that too?

 

[tabascosauz]

Ok, so the heck do I do that? You would think a brand new completely original unaltered cpu would not have to be messed with. I have been team red for a very long time but intel and Nvidia are looking so much better every day.

Mmmm I mean you could always run it stock, but both camps (Intel/AMD) kinda require you to at least do a bit of work if you want to get the most out of your CPU. Can't argue on the Nvidia front though, that is mostly true.

Negative/dynamic Vcore offset should be pretty self-explanatory and usually up front and easily visible on the main tweaker page. Been a while since I touched a MSI BIOS. You should leave it until after you've decided on a Curve Optimizer offset and are reasonably confident it's stable.

There are a lot of guides on Curve Optimizer, I don't think I can really explain it adequately. It's just an undervolting feature that assigns a number corresponding to a non-fixed millivolt value. 0 is stock, -30 is the max undervolt amount, +30 is the max overvolt amount (no one does or should really use CO to overvolt).

You'll find it under the name Kombo Strike in MSI BIOS, because initially (up until AGESA 1208) AMD globally removed CO control from the 5800X3D. In newer BIOSes iirc they might also have added proper Curve Optimizer back somewhere under Tweaker page or under PBO (Precision Boost Overdrive). Kombo Strike should have 3 settings, Kombo 1/2/3 corresponding to -10, -20 and -30 respectively.

In any case, I think MSI BIOS has a search function somewhere.

I noticed your gpu and cpu are smooth lines. Why is mine bouncing around like that for my cpu and gpu? Everyone else has smooth lines like that too?

CPU clock is not usually supposed to be smooth lines for Ryzen. But 5800X3D is a bit of a special CPU, it doesn't quite have the same boost behaviour as any other AM4 Ryzen. It is limited by a rather low inviolable Fmax ceiling (unless you BCLK) at either 4450 for multi core or 4550 for single core, so with the help of CO undervolting the CPU ends up "pegged" at the 4450 purple line because it's easily achievable, but is not allowed to go any higher.

What I know for sure is that you should never have major freq drops like in your Timespy run. Here's 5700G, no longer a smooth line at Fmax (due to limited cooling and a much higher 4850 Fmax), but still no major drops:

They used to run reasonably well out of the box, unfortunately, poor ol 5800X3D got the short end of the stick sometime in 2023(?). No one really knows exactly why there was such a big change to the V-F curves. Perhaps there were a lot of 5800X3Ds that were erroring out/crashing because the VIDs were too optimistic for the yields at the time, or something.

 

[Vya Domus]

27C delta at full load - that block should really be remounted.

That delta is never going to go down by much no matter how many times he remounts the block unless he uses PTM7950 or LM. The hotspot temp is absolutely fine for a water cooled 7900XTX, this isn't the problem.

 

[KedsDead]

Here is the results from running R23

I took some metrics for you all to look over.

Some taken while running Uningine With the 7900xtx OC in adrenalin (see settings pic) the highest gpu clock speed i saw was 2611. Highest hotspot was 66c. total board power average was 345-355 with a high spike of 525. The gpu memory clock speed seemed to hold steady at 2487.

Some specs are from running cpu-z benchmark

Here is also a link to the cpu-z benchmarks results https://valid.x86.fr/zhfc2k

When you took this screen shot.. What did you change?

Apply settings doesnt light up unless you changed something.

Power seems to low for what is being showed here.

My 7900xt runs more voltage and much higher clocks. ( Not the idle voltage, the voltage under load from his post )


This Screen doesnt match the one he showed as configured. or setting for the card.

Memory clock below factory speed, he says its clocked at 2700?

Voltage also seems way off from 1150mv

 

[LabRat 891]

Dunno if related to OP's issue but...

My 7900 GRE was underperforming and stuttering in some games, pretty badly.
Turns out, even PCIe 8-pins have a Insertion-removal-reinsertion lifespan, too.
Swapping to newer/less used PCIe power leads, resolved my issue.

 

[DonnieDangerous]

Mmmm I mean you could always run it stock, but both camps (Intel/AMD) kinda require you to at least do a bit of work if you want to get the most out of your CPU. Can't argue on the Nvidia front though, that is mostly true.

Negative/dynamic Vcore offset should be pretty self-explanatory and usually up front and easily visible on the main tweaker page. Been a while since I touched a MSI BIOS. You should leave it until after you've decided on a Curve Optimizer offset and are reasonably confident it's stable.

There are a lot of guides on Curve Optimizer, I don't think I can really explain it adequately. It's just an undervolting feature that assigns a number corresponding to a non-fixed millivolt value. 0 is stock, -30 is the max undervolt amount, +30 is the max overvolt amount (no one does or should really use CO to overvolt).

You'll find it under the name Kombo Strike in MSI BIOS, because initially (up until AGESA 1208) AMD globally removed CO control from the 5800X3D. In newer BIOSes iirc they might also have added proper Curve Optimizer back somewhere under Tweaker page or under PBO (Precision Boost Overdrive). Kombo Strike should have 3 settings, Kombo 1/2/3 corresponding to -10, -20 and -30 respectively.

In any case, I think MSI BIOS has a search function somewhere.



CPU clock is not usually supposed to be smooth lines for Ryzen. But 5800X3D is a bit of a special CPU, it doesn't quite have the same boost behaviour as any other AM4 Ryzen. It is limited by a rather low inviolable Fmax ceiling (unless you BCLK) at either 4450 for multi core or 4550 for single core, so with the help of CO undervolting the CPU ends up "pegged" at the 4450 purple line because it's easily achievable, but is not allowed to go any higher.

What I know for sure is that you should never have major freq drops like in your Timespy run. Here's 5700G, no longer a smooth line at Fmax (due to limited cooling and a much higher 4850 Fmax), but still no major drops:

View attachment 345726

They used to run reasonably well out of the box, unfortunately, poor ol 5800X3D got the short end of the stick sometime in 2023(?). No one really knows exactly why there was such a big change to the V-F curves. Perhaps there were a lot of 5800X3Ds that were erroring out/crashing because the VIDs were too optimistic for the yields at the time, or something.

Here are a few more pics with the gpu in default settings and OC through Adrenalin with the hotspots tab open

Here are a few more pics with the gpu in default settings and OC through Adrenalin with the hotspots tab open

running r23 now with HWinfo

Mmmm I mean you could always run it stock, but both camps (Intel/AMD) kinda require you to at least do a bit of work if you want to get the most out of your CPU. Can't argue on the Nvidia front though, that is mostly true.

Negative/dynamic Vcore offset should be pretty self-explanatory and usually up front and easily visible on the main tweaker page. Been a while since I touched a MSI BIOS. You should leave it until after you've decided on a Curve Optimizer offset and are reasonably confident it's stable.

There are a lot of guides on Curve Optimizer, I don't think I can really explain it adequately. It's just an undervolting feature that assigns a number corresponding to a non-fixed millivolt value. 0 is stock, -30 is the max undervolt amount, +30 is the max overvolt amount (no one does or should really use CO to overvolt).

You'll find it under the name Kombo Strike in MSI BIOS, because initially (up until AGESA 1208) AMD globally removed CO control from the 5800X3D. In newer BIOSes iirc they might also have added proper Curve Optimizer back somewhere under Tweaker page or under PBO (Precision Boost Overdrive). Kombo Strike should have 3 settings, Kombo 1/2/3 corresponding to -10, -20 and -30 respectively.

In any case, I think MSI BIOS has a search function somewhere.



CPU clock is not usually supposed to be smooth lines for Ryzen. But 5800X3D is a bit of a special CPU, it doesn't quite have the same boost behaviour as any other AM4 Ryzen. It is limited by a rather low inviolable Fmax ceiling (unless you BCLK) at either 4450 for multi core or 4550 for single core, so with the help of CO undervolting the CPU ends up "pegged" at the 4450 purple line because it's easily achievable, but is not allowed to go any higher.

What I know for sure is that you should never have major freq drops like in your Timespy run. Here's 5700G, no longer a smooth line at Fmax (due to limited cooling and a much higher 4850 Fmax), but still no major drops:

View attachment 345726

They used to run reasonably well out of the box, unfortunately, poor ol 5800X3D got the short end of the stick sometime in 2023(?). No one really knows exactly why there was such a big change to the V-F curves. Perhaps there were a lot of 5800X3Ds that were erroring out/crashing because the VIDs were too optimistic for the yields at the time, or something.

 

[MrDweezil]

Doesn't the "same hardware" chart on the Time Spy results show his system is basically running as it should?

 

[Redwoodz]

All I see is low cpu utilization and low cpu speeds. 24% and 3.6GHz. Amperage of 34w.

 

[3x0]

All I see is low cpu utilization and low cpu speeds. 24% and 3.6GHz. Amperage of 34w.

He didn't take a picture while CB is running. CPU package power max is 130+W.

The only thing I see abnormal is power reporting deviation is at ~110% min, presumably while running CB. That value should be closer to 100%

 

[Steevo]

Lower your Vmem speed back to default.

Lower the Mv in steps of 50 until it crashes, then by 10 or the closest step. The lower the Mv (voltage) allows for higher current at a set frequency/thermal load. My 7900XTX will run at 950Mv but is slightly unstable in games, so it sits around 975-1000Mv and it's very temp dependant. I know it sounds backwards compared to how we used to do it, but it's the same as using a negative offset for the CPU, and it truly is the silicon lottery.

The 7900Xtx is/was power limited, it's why I went aftermarket, when I get time again I am going to put mine on water and I hope they have BIOS mods to raise the power limit even higher.

 

[Redwoodz]

He didn't take a picture while CB is running. CPU package power max is 130+W.

The only thing I see abnormal is power reporting deviation is at ~110% min, presumably while running CB. That value should be closer to 100%

Well, hard to say what is what but I mentioned it for a reason. It's the only thing consistently lacking in his posts.

 

[DonnieDangerous]

Well, hard to say what is what but I mentioned it for a reason. It's the only thing consistently lacking in his posts.

Something missing from the testing you need to see that I didn't provide? I understand my cpu is underperforming a little and will look into how I can address that from the feedback provided in earlier posts but that doesn't explain why my card is performing badly? When first got it, my issues were all driver related but when it worked, it worked! Scores were always in the average range regardless of what benchmark I used. Now, only timespy is in the normal range. Kombuster, Uningine, and passmark all in the 70-75% range.

 

[Steevo]

Has gaming performance diminished or just benchmarks?

 

[DonnieDangerous]

Has gaming performance diminished or just benchmarks?

Hard to say because even with the lower benchmarks the 7900xtx is still a beast. None of the games I play have ever pushed it to a point where it is slowing down. I play farcry 6, starfield, fallout 4, division 2 all on max settings in 2560x1440. I play cities skylines 2 on my other monitor in 3840x2160 on max settings but that one usually gets throttled by the cpu

 

[Vya Domus]

If you are willing to I recommend a clean install of windows, just have it on a different partition and install nothing else besides the drivers and the superposition benchmark. Realistically this is the only way you're gonna conclusively determine if there is something wrong with the hardware.

 

[tabascosauz]

Here are a few more pics with the gpu in default settings and OC through Adrenalin with the hotspots tab open

running r23 now with HWinfo

So far you've showed us only OC results, right? What do your bone stock GPU results look like?

Because Wattman (in Radeon Software) isn't quite intuitive with the way it handles clocks and volts - what you set isn't really what you get. Maxing out Vcore slider might not result in the gains you expect. A bit like undervolting Nvidia these days (what you set isn't what you get in terms of clocks), but AMD complicates things further by varying Vcore as well.

As to the CPU side, things are pretty predictable. You aren't hitting 90C but you are almost there. Way too much power and thermals for too low clock, it's just post-1206 AGESA in action. Same solution - try to go as low as possible stable on CO, then start experimenting with negative Vcore offset.

If you are willing to I recommend a clean install of windows, just have it on a different partition and install nothing else besides the drivers and the superposition benchmark. Realistically this is the only way you're gonna conclusively determine if there is something wrong with the hardware.

Think the lad has said they've already clean installed a dozen times.

 

[DonnieDangerous]

So far you've showed us only OC results, right? What do your bone stock GPU results look like?

Because Wattman (in Radeon Software) isn't quite intuitive with the way it handles clocks and volts - what you set isn't really what you get. Maxing out Vcore slider might not result in the gains you expect. A bit like undervolting Nvidia these days (what you set isn't what you get in terms of clocks), but AMD complicates things further by varying Vcore as well.

As to the CPU side, things are pretty predictable. You aren't hitting 90C but you are almost there. Way too much power and thermals for too low clock, it's just post-1206 AGESA in action. Same solution - try to go as low as possible stable on CO, then start experimenting with negative Vcore offset.



Think the lad has said they've already clean installed a dozen times.

I have showed both multiple times. If you look at the names of the pictures, it will say something like Default. If OC it says OC. most of the pics are on default settings. I did a few runs in OC to show not much of a difference. For the cpu every pic is completely default settings. I have never touched the cpu. 99% my gpu is left on default.

 

[Vya Domus]

Post a 1440p/4K benchmark of furmark 2, that's a benchmark that does not put any significant load on the CPU, we'll know for sure if the CPU is the problem. If it's still underperforming then it's something else.

 

[tabascosauz]

I have showed both multiple times. If you look at the names of the pictures, it will say something like Default. If OC it says OC. most of the pics are on default settings. I did a few runs in OC to show not much of a difference. For the cpu every pic is completely default settings. I have never touched the cpu. 99% my gpu is left on default.

Is the "before R23" pic supposed to represent idle state? Because 40W idle is anything but, it's not even close to idle. 15W SOC power is whatever and will elevate your minimum idle power to about 25-30W, but 10W core power continuous points to a LOT of background crap or other programs you haven't closed.

When running either these CPU or GPU benchmarks, close as much stuff as you possibly can. It's not helping your scores, even if you said you clean installed multiple times, you may as well not have if you just reinstall the same heap of stuff and leave all that background stuff open (ie. Corsair service, whatever in task manager was handling the mini display).

Don't get too hung up on percentile metrics - find out just how MUCH lower the score is. Superposition is hard to compare because everyone runs different settings. The Timespy scores look pretty normal for an overall that's being weighed down by such a low CPU score. Passmark is so nebulous, it's not even worth talking about.

 

[Zach_01]

Run a superposition benchmark earlier but I was in a hurry and didn’t get screenshots. Not on my PC now.
Score under 4K optimized settings was around 24k.

R9 5900X
Nitro+ 7900XTX (390W TBP)
2x16GB 3600MT

I’ll be happy to run any benchmark for comparison and share any settings GPU drivers or any other.
I’ll share GPU operation conditions through HWiNFO64.

Later…

EDIT:
Mistyping 8K instead of 4K

 

[DonnieDangerous]

here is the results from lowering Mv. I was able to get a stable benchmark at 1040. A big improvement but still not around the 30k mark where it should be for a refernece card.

 

[Zach_01]

here is the results from lowering Mv. I was able to get a stable benchmark at 1040. A big improvement but still not around the 30k mark where it should be for a refernece card.

30k points at 4K optimized settings?

At unigine website there is a leaderboard and the best score of 7900XTX is 28.1k. Second best is 27.3k. Third is 26.6k. 4th scored 26.4k and so on…

 

[DonnieDangerous]

30k points at 4K optimized settings?

At unigine website there is a leaderboard and the best score of 7900XTX is 28.1k. Second best is 27.3k. Third is 26.6k
Rest of them below 26k

My bad, 30k is for timespy and passmark

Beginning to think it's a one off though. I can't get a benchmark over 20k now using same settings. Doing a 4th run now using afterburner instead of adrenalin

Is the "before R23" pic supposed to represent idle state? Because 40W idle is anything but, it's not even close to idle. 15W SOC power is whatever and will elevate your minimum idle power to about 25-30W, but 10W core power continuous points to a LOT of background crap or other programs you haven't closed.

When running either these CPU or GPU benchmarks, close as much stuff as you possibly can. It's not helping your scores, even if you said you clean installed multiple times, you may as well not have if you just reinstall the same heap of stuff and leave all that background stuff open (ie. Corsair service, whatever in task manager was handling the mini display).

Don't get too hung up on percentile metrics - find out just how MUCH lower the score is. Superposition is hard to compare because everyone runs different settings. The Timespy scores look pretty normal for an overall that's being weighed down by such a low CPU score. Passmark is so nebulous, it's not even worth talking about.

Yes, at idle. I task ended any extra programs but didn't really make much of a difference. The only thing running on tests I will be posting in a minute is HWinfo and either afterburner or adrenalin.

Is this a good tutorial for undervolting the 5800x3d or could you share a video of what I need to do please

 

[tabascosauz]

Yes, at idle. I task ended any extra programs but didn't really make much of a difference. The only thing running on tests I will be posting in a minute is HWinfo and either afterburner or adrenalin.

Is this a good tutorial for undervolting the 5800x3d or could you share a video of what I need to do please

Yeah then there is still something hogging all that CPU utilization then. I'm only ever at 10-15W core power if I'm say, running some light games and streaming in discord at the same time, even just web browsing is less than 5W core power. For sure that 10W should be having an impact on CPU perf in benches.

Looks like they have the right idea. You don't need to touch power limits, it's mostly just for seriously cooling limited setups. Just worry about CO, then Vcore offset.

Their mistake seems to be that they are also on a post-1206 BIOS and are fixating on PBO and CO. Which is fine if they were on earlier AGESA, but the new V-F changes can be drastic and the only way (after dialing in CO) to try and pull it back to normal is to use Vcore offset.