RX 9000 series GPU Owners Club

[Tomorrow]

I may later today, but you can see if you have the Triple Buffering checkbox turned on in AMD Adrenalin Control Panel > Gaming > Graphics, some games may interpret it as "Vsync".
This has probably been a bug for a while.

This setting only affects OpenGL from what i read and i dont even have it enabled. Just today i discovered that in one old game that uses DX11 (was DX9 initially) disabling the "optimizations for windowed games" windows setting restored performance i had lost when moving to AMD from Nvidia.
I suspect the whole <~200 fps thing is likely another such setting, somewhere. It's just odd that these three legacy versions are affected, but not much else.

 

[Bloste]

Sure, but interpreting something that is 0.00000000001ms as something <0 is like really badly incorrect.
If there were 0 ms, I'd say it's fine. But -1ms? Frame time can't be of negative value. Doesn't matter, but nice bug.

Well, it could be that 0.00000000001 < 0 for big enough values of 0.

Ok, ok, I know where the exit is. I'll see myself out.

 

[AVATARAT]

This setting only affects OpenGL from what i read and i dont even have it enabled. Just today i discovered that in one old game that uses DX11 (was DX9 initially) disabling the "optimizations for windowed games" windows setting restored performance i had lost when moving to AMD from Nvidia.
I suspect the whole <~200 fps thing is likely another such setting, somewhere. It's just odd that these three legacy versions are affected, but not much else.

I discovered this a while ago when I was testing something, not sure which game it was in, it was just weird because I tried turning vsync off and it finally stopped syncing after I turned TB off.

I also turn off those optimizations and manually set high values on the games/apps I care about below in that menu.

And yes, these are probably Microsoft's "green/energy saving" bs, I just turn off most of them that I know of, of course.

 

[Tomorrow]

Ok so i've made some progress. Apparently when i connect my monitor to my 9070 XT over HDMI there is no limit. Framerates are as expected in the four digits.
So whatever this is has something to do with DisplayPort specifically. I already tried both DP ports on my card. What i have not yet tested is alternate DP cable to connect my GPU and monitor. Currently im using the cable that came with the monitor.

EDIT: Ok i have solved the issue. For some unknown reason it was the fault of the DP cable that came with the monitor.
Using my own (previous cable) i was able to get proper framerates.

This is very bizarre. It's the first time i've had such an issue where a physical display cable is somehow limiting performance in some specific apps.
Normally if a DP cable limits anything it's the resolution, framerate, colors etc. Not fps in some apps only. Very strange.
I will have to do some more digging on this.

In the meantime when someone else buys this specific LG monitor then make sure you use your own cable: LG UltraGear 27GS85Q-B.
Naturally i only have sample size of one and most people do not run these legacy benchmarks, but it's possible that other cables included with LG monitors may be affected too. Especially if they use the same supplier/design.

EDIT 2: This keeps getting more and more bizarre. Apparently changing the cable set windows to duplicate mode (i have my TV connected via HDMI) which incorrectly led me to conclude it was the cable's fault.

When i disabled the duplicate mode and set it to PC Screen only (Win+P shortcut) the 200fps limit was back. Setting it to duplicate again removed it.
For some reason my Windows does not seems to allow full performance when i have it set to only one screen.

For some reason duplicate mode gives better performance. Now i will have to retest other benchmarks and games to see if this is universal or limited to these handful of legacy benchmarks.

 

[QuietBob]

@AVATARAT
OK, so I tested a full stock MBA 7900xtx in 3DMark 01SE, 03 and 06 using default settings. 01 shows an error on launch, and 06 during installation, but both run fine



These are my results, 01SE and 03 is 768p, and 06 is 1024p:



The GPU didn't even boost properly in such low resolutions. Though your results in 01SE and 06 are way higher, did you enforce a minimum clock in Adrenalin perchance?

 

[AVATARAT]

@AVATARAT
OK, so I tested a full stock MBA 7900xtx in 3DMark 01SE, 03 and 06 using default settings. 01 shows an error on launch, and 06 during installation, but both run fine

View attachment 400677View attachment 400678

These are my results, 01SE and 03 is 768p, and 06 is 1024p:

View attachment 400674View attachment 400675View attachment 400676

The GPU didn't even boost properly in such low resolutions. Though your results in 01SE and 06 are way higher, did you enforce a minimum clock in Adrenalin perchance?

My scores on 01SE and 06 are higher because I ran them with the default 3DMark settings.
On 03 I used custom (high) settings because I didn't want to see 10k FPS again That's why you have higher scores there, I can run it on default too if you want to compare them. Overall, I ran them just to see if they work, so the results might be a bit high in all the benchmarks

From my experience the 9070 XT is faster at lower resolutions than the 7900 XTX, which could be faster at higher resolutions.
And the clock management of the 9070 XT is different than the 7900 XTX, I mean we can't set a minimum clock like you can, we can only set the max (wished) boost clock + minus offset voltage, so in the end we have no control over the frequency.
I really liked the Vega control where we had control over each of the 7 states, but those days are gone

 

[QuietBob]

My scores on 01SE and 06 are higher because I ran them with the default 3DMark settings.
On 03 I used custom (high) settings because I didn't want to see 10k FPS again That's why you have higher scores there, I can run it on default too if you want to compare them. Overall, I ran them just to see if they work, so the results might be a bit high in all the benchmarks

From my experience the 9070 XT is faster at lower resolutions than the 7900 XTX, which could be faster at higher resolutions.
And the clock management of the 9070 XT is different than the 7900 XTX, I mean we can't set a minimum clock like you can, we can only set the max (wished) boost clock + minus offset voltage, so in the end we have no control over the frequency.
I really liked the Vega control where we had control over each of the 7 states, but those days are gone

Thanks for these insights, I didn't know the clock boosting mechanism works differently on RDNA4. Perhaps that would explain much higher scores in 3DMark01 and 06 in your case. I ran all the benchmarks on default settings, yet you scored 50% higher for most tests, and literally double in some

I know we're dealing with ST limited APIs and low resolutions, where your Ryzen 7 9700X has an advantage. Yes, we're effectively benchmarking CPUs here. But the architectural differences between Zen 5 and 3 (my CPU) aren't the only reason for the big discrepancy in the results IMO. It could be the way RDNA4 manages boost states.

For a sanity check, could you please run 3DMark again with these settings on a stock card, including feature/synthetic tests for each benchmark?

01SE - 4K + 8 sample AA
03 - 4K + 8 sample AA + 16x AF + postprocessing
06 - 4K + 8 sample AA + 16x AF + force full precision

I'm really curious how these two cards will stack up when the stress is shifted towards the GPU. Here are the screens for reference, thanks again!

 

[Bloste]

I have a weird question, sorry if it's silly: could it be possible that Samsung memory is not ECC mem? Because when I overclock the memory it doesn't reach that stage in which the performance decreases, but it goes directly to crash, or even hang Windows.

 

[spacer00ster]

Hey guys. Does Speed Way crash for anyone here? It crashes for me EVERYTIME - It loads, I hear what I believe is the benchmark audio (no video), then poof, it crashes to desktop and the driver recovers.

Note that I am on Unraid with GPU passthrough to Windows 11 VM. This weekend I plan to shut down Unraid and boot into a test Win11 disk I have but just wanted to know if anyone has ever experienced issues like this.

 

[Kwadratowicz]

@spacer00ster
I will try it later on my pc. (Win 10 ltsc)
@Bloste
i think all 9070 has ecc, but are You tuning only memory, or other settings are tuned as well?

 

[AusWolf]

Hey guys. Does Speed Way crash for anyone here? It crashes for me EVERYTIME - It loads, I hear what I believe is the benchmark audio (no video), then poof, it crashes to desktop and the driver recovers.

Note that I am on Unraid with GPU passthrough to Windows 11 VM. This weekend I plan to shut down Unraid and boot into a test Win11 disk I have but just wanted to know if anyone has ever experienced issues like this.

I only ran Speed Way once just to see if it works (I'm mainly on Linux, 3DMark doesn't work here), but had no problem. Maybe your VM setup is key to the issue?

 

[LittleBro]

I have a weird question, sorry if it's silly: could it be possible that Samsung memory is not ECC mem? Because when I overclock the memory it doesn't reach that stage in which the performance decreases, but it goes directly to crash, or even hang Windows.

 

[AusWolf]

View attachment 400749

May? I'd rather say it does.

Most 9070 cards use Hynix chips (all of the early batches), only a few manufacturers started mixing Samsung into the picture recently.

 

[Kwadratowicz]

And another rabbit hole ,

NVIDIA GeForce RTX 3080 Founders Edition Review - Must-Have for 4K Gamers

NVIDIA's new GeForce RTX 3080 "Ampere" Founders Edition is a truly impressive graphics card. It not only looks fantastic, performance is also better than even the RTX 2080 Ti. In our RTX 3080 Founders Edition review, we're also taking a close look at the new cooler, which runs quietly without...

www.techpowerup.com

for example gddr6x in rtx 3080 FE does not have ECC it has error correction that is similar. So it could be that ECC is not in the card, but something similar should be.

 

[spacer00ster]

@spacer00ster
I will try it later on my pc. (Win 10 ltsc)
@Bloste
i think all 9070 has ecc, but are You tuning only memory, or other settings are tuned as well?

Kool

I only ran Speed Way once just to see if it works (I'm mainly on Linux, 3DMark doesn't work here), but had no problem. Maybe your VM setup is key to the issue?

I'm thinking that it's a mix of Unraid and the VM. These new cards aren't officially supported on Unraid yet so I can live with Speed Way not working.

 

[LittleBro]

I can't find any information on whether H56G42AS8DX-014 supports ECC.

 

[AusWolf]

I can't find any information on whether H56G42AS8DX-014 supports ECC.

AMD has been using ECC memory on GPUs since RDNA 2 (or maybe RDNA 1). That's why sometimes a higher frequency results in lower performance instead of crashes, because ECC kicks in to correct the errors.

 

[Kwadratowicz]

Kool

It's working with no problems,
@Bloste
And because of Your comment, now I'm trying to overclock memory

 

[AVATARAT]

It's working with no problems,
@Bloste
And because of Your comment, now I'm trying to overclock memory

Just make sure it's not running a game because you can't change the VRAM frequency when playing a game, your entire system will freeze/restart

 

[Kwadratowicz]

Just make sure it's not running a game because you can't change the VRAM frequency when playing a game, your entire system will freeze/restart

Don't worry I've already checked that with power limit .
And for now 2800 vram was going great with speedway, around 7k, but 2850 was instant crash.

 

[LittleBro]

I can switch between my custom profile (which includes memory OC) and default one while gaming. It makes very brief blink to black and nothing crashes or stops working, game continues. Asus GPU Tweak 3

 

[AVATARAT]

Thanks for these insights, I didn't know the clock boosting mechanism works differently on RDNA4. Perhaps that would explain much higher scores in 3DMark01 and 06 in your case. I ran all the benchmarks on default settings, yet you scored 50% higher for most tests, and literally double in some

I know we're dealing with an ST limited API here and low resolutions, where your Ryzen 7 9700X has an advantage. Yes, we're effectively benchmarking CPUs here. But the architectural differences between Zen 5 and 3 (my CPU) aren't the only reason for the big discrepancy in the results IMO. It could be the way RDNA4 manages boost states.

For a sanity check, could you please run 3DMark again with these settings on a stock card, including feature/synthetic tests for each benchmark?

01SE - 4K + 8 sample AA
03 - 4K + 8 sample AA + 16x AF + postprocessing
06 - 4K + 8 sample AA + 16x AF + force full precision

I'm really curious how these two cards will stack up when the stress is shifted towards the GPU. Here are the screens for reference, thanks again!

View attachment 400736View attachment 400737View attachment 400738

OK - 01, 03, 06 Default settings for OC and Graphics.

I scored 0 in 3DMark03 Custom

AMD Ryzen 7 9700X, AMD Radeon RX 9070 XT x 1, 65536 MB, 64-bit Windows 11}

www.3dmark.com

I scored 0 in 3DMark06 Custom

AMD Ryzen 7 9700X, AMD Radeon RX 9070 XT x 1, 65536 MB, 64-bit Windows 11}

www.3dmark.com

Much appreciated mate. Now i know it's not a universal thing and the problem is somewhere in my PC. Hopefully it's software and i can get it fixed by changing some setting i have overlooked. If it's not too much trouble can you provide screenshots of the following settings:

Windows > System > Display > Graphics. Make sure advanced graphics settings options are also visible. Apps list it not necessary.
AMD Adrenalin Control Panel > Gaming > Graphics. I want to see what profile is in use and what Graphics Options list shows.
AMD Adrenalin Control Panel > Cogwheel (Settings) > Display. Make sure Display Specs and Overrides options are also visible.

Sorry for the late response, here they are:

 

[QuietBob]

OK - 01, 03, 06 Default settings for OC and Graphics.

Thanks a lot for your time! Look how the tables have turned now. Maxed out 4K (as detailed in my previous post), with the card running stock:



Sorry for the stitched shots. Win10 applied some compatibility settings, resulting in messed up UI. It would seem that the 7900xtx does better in legacy APIs. Of course, this finding is mostly of academic value with the number of fps either GPU churns out.

May I still ask for those feature tests in 06? Pretty please?

 

[Kwadratowicz]

06, why it looks a lot better than newest benchmarks? new benchmarks are like soulless and surgically crafted, not a speck of dust there, not like in old ones .



@QuietBob
Edit features tests :

 

[QuietBob]

Edit features tests :

Thanks!
I'd still like to see those in maxed out 4K for a complete GPU comparison.