AMD RX 7000 series GPU Owners' Club

[QuietBob]

Kinda hard to draw much power with a 60fps cap, though.

I prefer the term "calculated risk"

The 500w figure I quoted is the highest total system power consumption I've seen when running uncapped benchmarks. Interestingly, it uses most power at 1080p, possibly because the CPU has to work harder and the card boosts higher, even over 3100 MHz effective clock

My 5800X3D with -30 CO hits 105w in CB R23. I doubt any game could push it further. If the reference 7900XTX has transients of 455w, I still should be below 650w in a worst-case scenario. The GX-650 I'm using has never shut down, which I reckon points more to the driver when a benchmark randomly exits.

Did you test your SFX PSU with the new certified cable alone?

 

[tabascosauz]

It is pretty stupid, considering that 6000 series does the exact opposite - adjust the VRAM clock dynamically on the desktop and during video playback, and use the maximum during games.

That's a good point. During Youtube playback VRAM clock is 909MHz, not full speed.

Been looking into changing the blanking settings in CRU to get VRAM to downclock but Win 11 is pain

Did you test your SFX PSU with the new certified cable alone?

Nah, too lazy to swap it in/out another 2 times lol

Still leaning towards a power issue for the crashing specifically, as I've still never crashed since the switch. Cable choice does seem to make a difference for Freesync performance, however.

 

[remekra]

Just tested the idle power draw on my XTX with 23.1 drivers and at desktop it draws around 30-70W, with memory clock going up to 900MHz. When playing a youtube video at 4k, usage goes to 60-90W and memory clocks go up to 1500MHz so it is adjusting them. I am running one screen (42C2 oled) at 4K.

Also regarding nvidia, I've had 3080FE before that one, can't remember the idle power draw but when running the desktop at HDR (which I don't normally do) it had issues with scrolling. For the first 1-2 seconds after you started scrolling any page in browser it would lag, like really badly and then it would go away if you kept on scrolling. Start again and it would lag again. When I looked at the clocks it was running them really low and when you started to scroll GPU boosted up to almost max and when you stop it would go to idle again, thats where the lag came from.

Only solution was to change the power setting in nvidia CP to prefer performance, but then the GPU would not even go to fan stop, as the idle temps went up.
Didn't happen when HDR was disabled so didn't bother to investigate further.
But the same doesn't happen on XTX, power draw stays the same regardless of HDR.

 

[QuietBob]

Just tested the idle power draw on my XTX with 23.1 drivers and at desktop it draws around 30-70W, with memory clock going up to 900MHz.

Weird, mine shows much lower figures This is on the desktop with the browser window minimized, on a single 4K60 monitor:

 

[kapone32]

Weird, mine shows much lower figures This is on the desktop with the browser window minimized, on a single 4K60 monitor:

View attachment 280707

Could be the refresh rate?

 

[QuietBob]

Could be the refresh rate?

Double the refresh rate, double the power? Might as well be

 

[remekra]

With just desktop and minimized browser it hovers around 30, but yeah it's 120Hz and also I'm not running it on stock, but with raised PL. However I also did UV, so in theory when in idle it should draw less or the same.

 

[tabascosauz]

Just tested the idle power draw on my XTX with 23.1 drivers and at desktop it draws around 30-70W, with memory clock going up to 900MHz. When playing a youtube video at 4k, usage goes to 60-90W and memory clocks go up to 1500MHz so it is adjusting them. I am running one screen (42C2 oled) at 4K.

Also regarding nvidia, I've had 3080FE before that one, can't remember the idle power draw but when running the desktop at HDR (which I don't normally do) it had issues with scrolling. For the first 1-2 seconds after you started scrolling any page in browser it would lag, like really badly and then it would go away if you kept on scrolling. Start again and it would lag again. When I looked at the clocks it was running them really low and when you started to scroll GPU boosted up to almost max and when you stop it would go to idle again, thats where the lag came from.

Only solution was to change the power setting in nvidia CP to prefer performance, but then the GPU would not even go to fan stop, as the idle temps went up.
Didn't happen when HDR was disabled so didn't bother to investigate further.
But the same doesn't happen on XTX, power draw stays the same regardless of HDR.

The problem is not that VRAM clock doesn't have various clock states, it's that in terms of PURE idle just concerning multi-monitor refresh rate, drivers believe there is no in-between. It is either sub-100MHz, or 2487MHz. Thus it feels like this is something AMD could easily solve in drivers.

But even if they could it would make little practical difference (ie. 40W vs 85W), neither would be able to sustain fan stop for any reasonable amount of time.

How much is idle *really*? 30-70W is a huge range and sounds like it wasn't actually idle. Otherwise 30W for 900MHz VRAM doesn't sound too bad.

As long as 30W isn't interfering with anything else (ie. fan stop), seems fine

Double the refresh rate, double the power? Might as well be

Your idle numbers look the same as mine, but 4K120 iirc can only be done properly (4:4:4 10-bit) over HDMI 2.1, it's quite a bit of bandwidth. 4K60 is roughly the same as 1440p120/144

 

[remekra]

As long as 30W isn't interfering with anything else (ie. fan stop), seems fine

Yeah, I'm not bothered by it anymore, with Gigabyte version fans are at 0 rpm when on desktop. And even when I had MBA, when they were running at 500 rpm they were unaudible.

Your idle numbers look the same as mine, but 4K120 iirc can only be done properly (4:4:4 10-bit) over HDMI 2.1, it's quite a bit of bandwidth. 4K60 is roughly the same as 1440p120/144

Yup, running 4:4:4 12bit over HDMI 2.1.

 

[tabascosauz]

AMD's drivers team is charging into 2023 with aspirations of becoming very unfunny comedians

Not only did 23.1.2 NOT do anything for multi-monitor power consumption, it made power draw HIGHER

• On 23.1.1 (old), although CRU and Radeon blanking tricks did not work for >60Hz, 60Hz on Monitor #2 still results in same idle power as single monitor setup.
• On 23.1.2 (new), VRAM clock is ALWAYS 2487MHz.

1 x 1440p @ 165Hz 4:4:4 8-bit Freesync ON + 1 x 1440p @ 60Hz 4:4:4 10-bit Freesync OFF

23.1.1:

23.1.2:

 

[Space Lynx]

AMD's drivers team is charging into 2023 with aspirations of becoming very unfunny comedians

Not only did 23.1.2 NOT do anything for multi-monitor power consumption, it made power draw HIGHER

• On 23.1.1 (old), although CRU and Radeon blanking tricks did not work for >60Hz, 60Hz on Monitor #2 still results in same idle power as single monitor setup.
• On 23.1.2 (new), VRAM clock is ALWAYS 2487MHz.

1 x 1440p @ 165Hz 4:4:4 8-bit Freesync ON + 1 x 1440p @ 60Hz 4:4:4 10-bit Freesync OFF

23.1.1:
View attachment 280784

23.1.2:
View attachment 280785

sorry to see this mate, but you have at least confirmed for me I am sticking with my 6800 xt.

 

[AusWolf]

AMD's drivers team is charging into 2023 with aspirations of becoming very unfunny comedians

Not only did 23.1.2 NOT do anything for multi-monitor power consumption, it made power draw HIGHER

• On 23.1.1 (old), although CRU and Radeon blanking tricks did not work for >60Hz, 60Hz on Monitor #2 still results in same idle power as single monitor setup.
• On 23.1.2 (new), VRAM clock is ALWAYS 2487MHz.

1 x 1440p @ 165Hz 4:4:4 8-bit Freesync ON + 1 x 1440p @ 60Hz 4:4:4 10-bit Freesync OFF

23.1.1:
View attachment 280784

23.1.2:
View attachment 280785

Well, at least they did... something.

 

[brutlern]

AMD's drivers team is charging into 2023 with aspirations of becoming very unfunny comedians

Not only did 23.1.2 NOT do anything for multi-monitor power consumption, it made power draw HIGHER

• On 23.1.1 (old), although CRU and Radeon blanking tricks did not work for >60Hz, 60Hz on Monitor #2 still results in same idle power as single monitor setup.
• On 23.1.2 (new), VRAM clock is ALWAYS 2487MHz.

1 x 1440p @ 165Hz 4:4:4 8-bit Freesync ON + 1 x 1440p @ 60Hz 4:4:4 10-bit Freesync OFF

I just installed the new drivers for my 7900XT. There is no change for me from the old driver. Idle power consumption 40-50W, memory clock stays at 909Mhz with single monitor 3440*1440 - 144Hz - HDR On

 

[Super Firm Tofu]

AMD's drivers team is charging into 2023 with aspirations of becoming very unfunny comedians

Not only did 23.1.2 NOT do anything for multi-monitor power consumption, it made power draw HIGHER

• On 23.1.1 (old), although CRU and Radeon blanking tricks did not work for >60Hz, 60Hz on Monitor #2 still results in same idle power as single monitor setup.
• On 23.1.2 (new), VRAM clock is ALWAYS 2487MHz.

1 x 1440p @ 165Hz 4:4:4 8-bit Freesync ON + 1 x 1440p @ 60Hz 4:4:4 10-bit Freesync OFF

23.1.1:
View attachment 280784

23.1.2:
View attachment 280785

If you're feeling bored, try uninstalling 23.1.2 and then reinstalling. I did that with various driver versions when dealing with idle memory clocks on RDNA2 and would encounter different results - one install never clocking down, and another clocking back down to where it should have been. Same driver package.

 

[tabascosauz]

single monitor

Kinda self explanatory eh

If you're feeling bored, try uninstalling 23.1.2 and then reinstalling. I did that with various driver versions when dealing with idle memory clocks on RDNA2 and would encounter different results - one install never clocking down, and another clocking back down to where it should have been. Same driver package.

In an effort to specifically prevent this I did clean install both the upgrade and downgrade. And rebooted a couple of times while I was reapplying all my settings. But I might do it again.

If it wasn't a one-off fluke I suspect AMD is also trying to address poor performance/black screen flickering/audio cutting out from VRAM downclocking in CPU-bound games. For example, logically SC2 and Genshin both now no longer suffer unnecessary hitching and stuttering due to VRAM falling below 2487MHz.

But completely sacrificing idle power is a really stupid way to go about solving that problem; idle is idle and load is load. If I want I can just achieve the same in 5 seconds on 23.1.1 by setting the other monitor to 165Hz......

 

[QuietBob]

In one of my posts I mentioned that FFXV benchmark would sometimes exit before finishing. And SOTTR benchmark would crash if I set RT to high or ultra (although the game itself could be played normally). I thought it may be related to the immature driver, @tabascosauz advised that the undersized power suply and/or uncertified DP cable may be the reason. Turns out it was none of these.

The system was running out of virtual memory I had the swap file limited to 4 GB, while SOTTR benchmark with high/ultra RT will allocate up to 8 GB. FFXV benchmark with max settings needs 6 GB. What's strange, total physical memory load peaked at around 7200 MB when running either of them.

Also, I can say quite confidently that a 650w Tier A PSU can run a reference 7900XTX with stock clocks without an issue. As long as your other system components use less than 200w under load, there should be enough room for GPU transient spikes.

 

[AusWolf]

In one of my posts I mentioned that FFXV benchmark would sometimes exit before finishing. And SOTTR benchmark would crash if I set RT to high or ultra (although the game itself could be played normally). I thought it may be related to the immature driver, @tabascosauz advised that the undersized power suply and/or uncertified DP cable may be the reason. Turns out it was none of these.

The system was running out of virtual memory I had the swap file limited to 4 GB, while SOTTR benchmark with high/ultra RT will allocate up to 8 GB. FFXV benchmark with max settings needs 6 GB. What's strange, total physical memory load peaked at around 7200 MB when running either of them.

Also, I can say quite confidently that a 650w Tier A PSU can run a reference 7900XTX with stock clocks without an issue. As long as your other system components use less than 200w under load, there should be enough room for GPU transient spikes.

Why some games still rely on virtual memory even when you have more than enough system RAM is beyond me.

 

[tabascosauz]

If you're feeling bored, try uninstalling 23.1.2 and then reinstalling. I did that with various driver versions when dealing with idle memory clocks on RDNA2 and would encounter different results - one install never clocking down, and another clocking back down to where it should have been. Same driver package.

Is no bueno yo. Shit's still F'ed, single monitor is fine but turn on the second one and it goes haywire

Looks like 165hz+60hz is the closest i'll get

 

[Super Firm Tofu]

Is no bueno yo. Shit's still F'ed, single monitor is fine but turn on the second one and it goes haywire

Looks like 165hz+60hz is the closest i'll get

View attachment 280862

Well that's a major bummer. Gives me a sad.

Going to stick with 23.1.2, or roll back? They eventually got single monitor high refresh on RDNA2 right. I think I was at 9-13w @ 1440p/165 by the end, where previously it was in the 30's and 40's at idle, or could be depending on how the stars aligned.

I suppose there's a chance the next one will fix it.

 

[tabascosauz]

Well that's a major bummer. Gives me a sad.

Going to stick with 23.1.2, or roll back? They eventually got single monitor high refresh on RDNA2 right. I think I was at 9-13w @ 1440p/165 by the end, where previously it was in the 30's and 40's at idle, or could be depending on how the stars aligned.

I suppose there's a chance the next one will fix it.

I still roll back to 23.1.1. 90W idle almost doubles my UPS power draw.

Single monitor is fine, usually see close to 15W idle at 165Hz. Anything below 40W still allows the MBA cooler to stop the fans. Makes sense since 23.1.1 downclocks the VRAM to the same speed on 1 monitor or 2. If cheap-ass Fabric-at-home (Ryzen IFOP) can't reach single digits, I doubt it's easy for 1TB/s Fabric-on-crack (Fanout link); both GPU and VRAM clock already idle very low.

Would be so easy to fix, just either downclock or open up those intermediate 900MHz clocks for multi monitor, and introduce a global or per-game Force Max Performance switch in Radeon Settings. But it's AMD software team so nope

 

[QuietBob]

Borderlands 3 full benchmark run at 4K max settings (no AA), uncapped and with 60Hz V-sync:

TBP down by 117w for 34% lower power draw on average. GPU temps 7c and 13c lower. The game stutters horribly with uncapped fps, that highest spike is 225ms

 

[moicestkarl]

Hello !
I'm Karl from France and I'm new here.
I hope that I'm on the right place to find some help.
I'm about to buy a 7900XTX. I hesitate between two different models.
The XFX SPEEDSTER MERC 310 for 1199€ or the Gigabyte GAMING OC 24G for 1155€
Which one would you choose ?
The XFX is 50€ more than the Gigabyte.

And I have a second question.
I have a 3 screen set-up. (2x4k 60hz Display port and 1x4k120hz HDMI)
Is it a problem to handle for the 7900 XTX ?
Thank you very much.

 

[tabascosauz]

Here's a new doozy - anyone confirm on their card for 23.1.1? Logging data taken over the span of 60 seconds each.

So the two issues of high video playback power + stuttering/in-game FPS fluctuations with video playback on secondary monitor seem to be caused by some sort of bug with the Youtube player when in its default size (small player). Make the video fullscreen and watch core/VRAM clock and power quickly settle down again.

Doesn't matter whether on main screen or secondary screen, the rule holds.

• Not sure if other video sources are the same yet (e.g. VLC).
• Not sure if single monitor is affected.
• Not sure if fixed in 23.1.2, I need to go back and check but I'm not sure if I can properly test since my VRAM clock is pegged at 2487 on 23.1.2.
• If core clock fluctuations are mirrored in 23.1.2, I'll bug report to AMD.

 

[Space Lynx]

I still roll back to 23.1.1. 90W idle almost doubles my UPS power draw.

Single monitor is fine, usually see close to 15W idle at 165Hz. Anything below 40W still allows the MBA cooler to stop the fans. Makes sense since 23.1.1 downclocks the VRAM to the same speed on 1 monitor or 2. If cheap-ass Fabric-at-home (Ryzen IFOP) can't reach single digits, I doubt it's easy for 1TB/s Fabric-on-crack (Fanout link); both GPU and VRAM clock already idle very low.

Would be so easy to fix, just either downclock or open up those intermediate 900MHz clocks for multi monitor, and introduce a global or per-game Force Max Performance switch in Radeon Settings. But it's AMD software team so nope

please report this bug through the AMD report tool, I know they are already aware of it somewhat, but maybe not to this extent. It could be helpful to them.

 

[QuietBob]

Here's a new doozy - anyone confirm on their card for 23.1.1? Logging data taken over the span of 60 seconds each.

So the two issues of high video playback power + stuttering/in-game FPS fluctuations with video playback on secondary monitor seem to be caused by some sort of bug with the Youtube player when in its default size (small player). Make the video fullscreen and watch core/VRAM clock and power quickly settle down again.

Doesn't matter whether on main screen or secondary screen, the rule holds.

• Not sure if other video sources are the same yet (e.g. VLC).
• Not sure if single monitor is affected.
• Not sure if fixed in 23.1.2, I need to go back and check but I'm not sure if I can properly test since my VRAM clock is pegged at 2487 on 23.1.2.
• If core clock fluctuations are mirrored in 23.1.2, I'll bug report to AMD.

View attachment 281021 View attachment 281022

Currently testing 23.1.2 on a single 4K60 monitor, so my results are not really comparable. Default player size:

720p
front end 1316-1330 MHz (140-229 effective), shader 1369-1389 MHz (131-216 effective), RAM 127-216 MHz, TBP 40-41 W

1080p
front end 1318-1324 MHz (141-188 effective), shader 1369-1390 MHz (134-178 effective), RAM 180-282 MHz, TBP 41-42 W

1440p
front end 1317-1323 MHz (132-185 effective), shader 1369-1391 MHz (126-176 effective), RAM 170-909 MHz, TBP 42-55 W

2160p
front end 1317-1339 MHz (134-298 effective), shader 1369-1393 MHz (127-281 effective), RAM 909 MHz flat, TBP 55-59 W

No difference from the previous driver. Front end and shader clocks are the same across all resolutions, so the increased power consumption is squarely due to elevated memory clocks. This was already evident in the launch review here, showing maximum VRAM clock during local 4K video playback: