RX 9000 series GPU Owners Club

[Chrispy_]

I had some time to do 28 quick runs of Steel Nomad, i.e. I was watching TV while running quick 1-minute tests.

The flatness of those results proves how far into diminishing returns we are with power efficiency - for my sample at least. A -15% power limit is entirely negated by a modest -75mv undervolt. FWIW I've settled on 250W (-17% power limit) at -70mv as a daily driver, I'm within 1% of default performance but 250W is a mental threshold I didn't want to cross as that's heat being dumped into the room I have to sit in when using the PC. The stablility threshold of my card seems to be -80mv to -85mv in the two tests I've found that cause crashes from undervolting where actual games are seemingly stable at -100mv

Also here's the chart with the y-axis cropped since it's more readable:

 

[Kwadratowicz]

One issue I observed on you case , the rear case fan is too high positioned to help with the GPU cooling but will help with the CPU.

I'm waiting for the arctic p14 max pwm, cause the top fans are under powered, about rear it is good in Witcher 3 with 50% fan set up it was blowing out really hot air, so I think it works, and in the top the hot air was going out, but with the stock case fans it was really not that great, I'm thinking too add the ducts for the air, cause in front the problem is that only bottom fan is adding air to graphics card, and I would have 2 fans specifically adjusted with duct that they are supplying gpu with cold air from front, then the last fan on front would be blowing out.

Your Air CPU cooler is well positioned for such setup for your GPU backplate.

I need to change the nvme radiator cause it is to high and with lower radiator I'm able to put second fan on cpu cooler that would get the place almost exactly where the pull fan for gpu could be, a little more heat for cpu won't be a problem with 2700, which is 65W stock, and cooler that is rated for 200W of power .

And I'm thinking of adding second duct that will be placed between side panel and plastic shroud of gpu, that the hot air won't be recirculated to gpu.

But I for now I will just add the fans, and watch if it will change the temperatures just with fans that are bigger and better .

And I agree with You for ditching the bottom fan, cause now this fan is only providing the cooling and air to the end of the card, and is disrupting the flow of intake fan, but I will test that if it show much difference with or without this fan first, cause why not . But I need to find better way to test than Witcher 3 on Cachyos cause I have hard time recording values.

Edit I will use gpuz with logs in windows for the tests, cause it is way easier for Me now to set this up , and dual boot is such a great thing to troubleshoot hardware, and software
Edit2 I've tested, and bottom fan is making that no air is going to the back of the card, where the IO is, so the bottom fan has to go.

 

[rusty caterpillar]

The reason we think we have Samsung memory is because gpu-z said so. It could be an error, but I'm quite reluctant to disassemble my baby to reveal the truth:

No need for you to tear down your card. Do you have any pictures tear down of 9000 showing Samsung chips?

TBH I would like some of 9000 to have Samsung and some SK Hynix but, is not the case. At least this first batches have all Hynix. Maybe are cheaper than Samsung or Micron or have tech specs suitable for 9000 PCB. The only tear downs showing Samsung chips are GDDR7 on 5070.

 

[Chrispy_]

No need for you to tear down your card. Do you have any pictures tear down of 9000 showing Samsung chips?

TBH I would like some of 9000 to have Samsung and some SK Hynix but, is not the case. At least this first batches have all Hynix. Maybe are cheaper than Samsung or Micron or have tech specs suitable for 9000 PCB. The only tear downs showing Samsung chips are GDDR7 on 5070.

Are there any tools that can independently read memory information like Thaiphoon for desktop RAM?

I'm not in a hurry to dismantle my brand new GPU, but I've been using GPU-Z for at least a decade and it's been reasonably accurate on VRAM for me, getting all my mining GPUs correct for both the bitcoin and ETH rushes. Most of my cards had vBIOSes with dual or triple-entry for the different memory manufacturers and I could verify which brand memory was in use because if I altered BIOS timings for the wrong brand, the settings wouldn't result in improved hashrates. GPU-Z so far has a 100% flawless track record for me, so whilst I don't presume it's perfect, it's certainly not failed me yet.

If it's any help, my GPU was a pre-order so presumably it's one of the very newest batches, and not the stock that's been gathering in retailer warehouses since Christmas last year, awaiting the March launch. My GPU is almost certainly 3-4 months newer than some of the launch day stock.

Edit:
So I'm fine-tuning my undervolt now with memory speeds and for whatever reason, I'm getting nothing out of VRAM overclocking in Steel Nomad.
2800Mhz fast timings are only 1.4% faster than 2518 default timings (stock). 2850 is actually slower, presumably because of memory errors and ECC.
This is unexpected, since most people seem to get significant gains out of their VRAM clocks. Then again, most people seem to have Hynix modules, not Samsung.

 

[3x0]

This is unexpected, since most people seem to get significant gains out of their VRAM clocks. Then again, most people seem to have Hynix modules, not Samsung.

Try other tests, not just Steel Nomad

 

[AVATARAT]

I had some time to do 28 quick runs of Steel Nomad, i.e. I was watching TV while running quick 1-minute tests.

View attachment 395220

The flatness of those results proves how far into diminishing returns we are with power efficiency - for my sample at least. A -15% power limit is entirely negated by a modest -75mv undervolt. FWIW I've settled on 250W (-17% power limit) at -70mv as a daily driver, I'm within 1% of default performance but 250W is a mental threshold I didn't want to cross as that's heat being dumped into the room I have to sit in when using the PC. The stablility threshold of my card seems to be -80mv to -85mv in the two tests I've found that cause crashes from undervolting where actual games are seemingly stable at -100mv

Also here's the chart with the y-axis cropped since it's more readable:
View attachment 395223

Lowering the frequency by 150-300 MHz can help your test be more precise. Since these chips have a target (and probably maximum) frequency of 3450MHz and if there is room, they will jump as close as possible.

 

[Chrispy_]

Try other tests, not just Steel Nomad

Minimal difference. Tried Port Royal, Time Spy Extreme, Heaven, Superposition, and CP2077's built-in benchmark using the path-tracing preset. I'm going to leave it on 2750/fast instead of the stock 2518/default, because 1% free performance is still free performance

Lowering the frequency by 150-300 MHz can help your test be more precise. Since these chips have a target (and probably maximum) frequency of 3450MHz and if there is room, they will jump as close as possible.

This has helped, thanks. I've set it to -450 which presumably caps my card at 3.0GHz exactly, but I'm not interested in ever consuming the 335W and high voltages that >3GHz speeds are going to require. I now seem to have a 90-minute (and still going) stable setup with Steel Nomad at -105mv and that was the most sensitive thing which crashed after about 15-20 minutes when running -85mv with an uncapped boost clock.

More importantly, that means I can set my daily-driver undervolt to -100mv which, in conjunction with memory at 2750/fast means I should be able able to meet stock performance while knocking another 15-20W off the power consumption. 304W to ~235W is a pretty solid result to me, I find it much more appealing than the ~6% performance gain I could get from those 70W or so.

 

[rusty caterpillar]

Are there any tools that can independently read memory information like Thaiphoon for desktop RAM?

I'm not in a hurry to dismantle my brand new GPU, but I've been using GPU-Z for at least a decade and it's been reasonably accurate on VRAM for me, getting all my mining GPUs correct for both the bitcoin and ETH rushes. Most of my cards had vBIOSes with dual or triple-entry for the different memory manufacturers and I could verify which brand memory was in use because if I altered BIOS timings for the wrong brand, the settings wouldn't result in improved hashrates. GPU-Z so far has a 100% flawless track record for me, so whilst I don't presume it's perfect, it's certainly not failed me yet.

If it's any help, my GPU was a pre-order so presumably it's one of the very newest batches, and not the stock that's been gathering in retailer warehouses since Christmas last year, awaiting the March launch. My GPU is almost certainly 3-4 months newer than some of the launch day stock.

Edit:
So I'm fine-tuning my undervolt now with memory speeds and for whatever reason, I'm getting nothing out of VRAM overclocking in Steel Nomad.
2800Mhz fast timings are only 1.4% faster than 2518 default timings (stock). 2850 is actually slower, presumably because of memory errors and ECC.
This is unexpected, since most people seem to get significant gains out of their VRAM clocks. Then again, most people seem to have Hynix modules, not Samsung.

I can't say about GPUZ accuracy I didn't use it extensive.

I'm not saying is not possible to have Samsung chips on 9000 series. Could be that the 2 manufacturers sliced the market in 2 GDDR7/ GDDR6 while Samsung focused on GDDR7, SK Hynix focused on GDDR6 I guess is cheaper to not fork your efforts and costs.
Till somebody will find Samsung under the hood of 9070 or 9070XT I'll stay on what I know for now, the existent tear downs.

 

[Krit]

Weight difference between Asus Prime RX 9070 XT and Sapphire Pure RX 7800 XT

 

[lexluthermiester]

Weight difference between Asus Prime RX 9070 XT and Sapphire Pure RX 7800 XT
View attachment 395368View attachment 395369

13grams is not a big difference. I like the look of the Sapphire pure more!

 

[Event Horizon]

Weight difference between Asus Prime RX 9070 XT and Sapphire Pure RX 7800 XT
View attachment 395368View attachment 395369

My 9070 Prime ended up being noticeably lighter than the 3060 Ti Gaming X Trio it replaced. Pretty neat.

 

[Noci]

I would love to join this select club of RX 9000 series owners, but the household finance CEO (my wife) is reluctant to funding my investments in new hardware.
Noticed that when I said; 'Hun give me € 1,000 because I have to buy some stuff for my pc' and she returned with a scary look in her eyes, holding a cast iron frying pan .

So for now I'll hold on to my trusted RX 6700XT which serves me well to it's abillities .

Here locally the prices are dropping, but still the cheapest RX 9070XT atm is approximately € 780 and me being the odd one, wants to watercool it (no matter how good the air cooler is).
So an additional € 200 is most likely needed.

That being said, the only (full cover) RX 9070XT waterblock up to now (Alphacool) was for the Asrock Taichi (€ 899 local price), but that, hopefully will change soon:

Alphacool is working on more full cover waterblocks for different types of RX 9070XT's .
Don't mind Eddy's typo, they don't stop production, just development on more types of RX 9070XT's.

So for me there is hope and probably also some of you if you like watercooling your GPU.

Here's a link to their thread: https://forum.alphacool.com/threads/what-do-you-want.61/page-11

Still hope to join this club in the near future

 

[rusty caterpillar]

13grams is not a big difference. I like the look of the Sapphire pure more!

Unfortunately Pure is 3 slot card on paper only.





Pros: shroud doesn't cover much the exhausts of the heatsink.

 

[Kwadratowicz]

Alphacool is working on more full cover waterblocks for different types of RX 9070XT's .

I'm thinking of making first time watercooling, cause the temps are still not there, but the card is really great

 

[kurosagi01]

The RX9070XT hype tempted me back to the desktop space. Just so happened personal circumstances worked in my favour.
XFX Swift 9070 XT Black for £689.99 from Scan UK and purchased on the 10th April 2025

 

[rusty caterpillar]

I'm thinking of making first time watercooling, cause the temps are still not there, but the card is really great

View attachment 395452

I would change the case before doing that, front fans has too much obstruction on the sides exactly the most efficient part of the fans are obstructed by the metal sheens, Noctua down there has the airflow obstructed by the PSU chamber, most of it.

If you don't have the funds you can do the cuttings maybe more elegantly than me however the Noctua remains more than a half obstructed.



Still for the card you have enough room for putting a fan on the back of the card, if you meant water cooling for CPU only, even more room.

 

[TSiAhmat]

I'm thinking of making first time watercooling, cause the temps are still not there, but the card is really great

tbh I am not sure if your top right case fan is even doing a thing or rather your temps might be better if it is off, have you tested it at low rpms or off?

 

[Kwadratowicz]

If you don't have the funds you can do the cuttings maybe more elegantly than me however the Noctua remains more than a half obstructed.

I'm from the past era, where gpu that was 180W was almost top tier, and after searching the temperatures in my gigabyte are great, I've got ~3200 mhz boost all the time, with 310-320W of power draw (stock), and even in one revie they've had 90C and clocks below 2800mhz.
About cutting it is not worth it, I've spent week trying to close this case with the case, and even with other case, for now i will just make a little tweaks to fans, and some time to rest from reinstalling fans and components.
I've thought of water cooling the card, but it is not worth it with the budget for this build that i have.

tbh I am not sure if your top right case fan is even doing a thing or rather your temps might be better if it is off, have you tested that at low rpms or off?

Top fans are connected to 1, and yeah top right fan is just throwing out cold air, that's why i will try with 2 front fans disabled, and only noctua on for inlet.

 

[TSiAhmat]

Top fans are connected to 1, and yeah top right fan is just throwing out cold air, that's why i will try with 2 front fans disabled, and only noctua on for inlet.

Have you tried placing a book on top of it, it might help (just to see if there is any change, but not perfect since it still sucks air)

 

[Kwadratowicz]

Have you tried placing a book on top of it, it might help (just to see if there is any change, but not perfect since it still sucks air)

I will just remove it from connector , everything is really quite easy to disconnect , but first I need to check if the middle and top front fans are needed.

 

[TSiAhmat]

I will just remove it from connector , everything is really quite easy to disconnect , but first I need to check if the middle and top front fans are needed.

i personally would first remove it, since it throws some of the cold air out of the system which the top fan provides, also you will have negative air pressure with one fan in & two/three fans out unless the intake has very high rpm

 

[NSR]

Not sure if this has been reported already but with all of my fans at maximum, with my case panels removed, no dust filters etc. VRAM temp will not drop below 50 degrees C. This is rather odd, no? I'm not one of those people who is super concerned about my VRAM temps but I thought I'd test how low I could get them and they will just not go below 50.

Case: Corsair 4000D Airflow, x6 1300RPM fans, x3 intake, x3 outtake.
GPU: 9070 NITRO+ with backplate removed, x3 3800 RPM fans.

 

[TSiAhmat]

Not sure if this has been reported already but with all of my fans at maximum, with my case panels removed, no dust filters etc. VRAM temp will not drop below 50 degrees C. This is rather odd, no? I'm not one of those people who is super concerned about my VRAM temps but I thought I'd test how low I could get them and they will just not go below 50.

Case: Corsair 4000D, x6 1300RPM fans, x3 intake, x3 outtake.
GPU: 9070 NITRO+ with backplate removed, x3 3800 RPM fans.

removing case panels could make airflow weird/worse, can you try it with them on?

also i would slightly decrease outtake rpm for slight positive pressure. (can help for good airflow & dust build up in some cases)

 

[droopyRO]

Those memory temps on this series are wild. Check mine, 72ºC with 700rpm on the cooler(zero fan disabled) while watching Youtube, and the vRAM is maxed out at 2500Mhz. Yes, i have Bitdefender and ran a scan, so no malware. In idle it drops the clocks to like 60Mhz but the temperature is still 70º or above:

 

[TSiAhmat]

Those memory temps on this series are wild. Check mine, 72ºC with 700rpm on the cooler(zero fan disabled) while watching Youtube, and the vRAM is maxed out at 2500Mhz. Yes, i have Bitdefender and ran a scan, so no malware. In idle it drops the clocks to like 60Mhz but the temperature is still 70º or above:

that sounds rather high, I have around 58 ~ 60 C° in the same scenario (with no rpm)