RX 9000 series GPU Owners Club

[biffzinker]

Is anyone else doing a 3,322/26 MHz boost clock on their 9070? It seems the GPU I have can take the bump from 2,700 to 3,322. Another 10% to the PL shows 269.5W

I did a short stress test and it passed.

 

[iSpeakVeryWell]

Hello there and welcome, fellow owners of a 9000 series card! Hope you're all enjoying the hell out of it!

I've added @kurosagi01, @cadaveca, @biffzinker, @Krit, @iSpeakVeryWell, and @Redwoodz to the list. Feel free to share any missing details you'd like added to the front-page, or let me know if you'd prefer to be removed.
Thanks for your support, and once again, enjoy your new beasts!

Here are the missing data:

Purchased: 2025-03-22
Store: canadacomputers.com
Price: 1069$ CAD
VAT: not included
Canada

 

[biffzinker]

Hello there and welcome, fellow owners of a 9000 series card! Hope you're all enjoying the hell out of it!

I've added @kurosagi01, @cadaveca, @biffzinker, @Krit, @iSpeakVeryWell, and @Redwoodz to the list. Feel free to share any missing details you'd like added to the front-page, or let me know if you'd prefer to be removed.
Thanks for your support, and once again, enjoy your new beasts!

My card is 9070 minus XT, and it’s from BestBuy.

 

[DrSteveBrule]

long time lurker, first time poster. ordered an Asrock Steel Legend 9070xt the other day from newegg. I wasn't specifically in search of a card, but the price was too good with all the uncertainty ($699+tax US). I've searched through the thread and aside from a Taichi mention, I don't believe there are any Steel Legends, so I can be a data point soon. UPS is being a jerk by pushing back the expected delivery twice and it may not be here until Thursday.

 

[Kwadratowicz]

For how long have you been playing games on linux?

New GPU new system , I was fed up with windows, and trying linux for around 3 weeks now.
I've even thrown everything down, and changed back to windows for half a day to think that linux is way more friendly for Me

 

[rusty caterpillar]

I don't agree with that, cause if You look at the gpu pcb, 9070 XT has a lot denser pcb with vram really close to the gpu core, but in older cards You've had vram farther from gpu core.
Here is gpu pcb from ASRock Radeon RX 9070 XT Taichi OC (from techpowerup review)
View attachment 396174
Here is gpu pcb from ASRock Radeon RX 7900 XTX Taichi (from techpowerup review)
View attachment 396175

I think that some of that 6C cooler vram in My case is from cooler gpu core. (I don't know if it is by chance or it is corelated, but the gpu core temp has gone down 6C like vram).

And the cooler for vram is also cooler for gpu or is connected to it (in most cases), so it should be corelated. (Corrected)
Edit 1 And I disabled front fan which is on top of the front, I will see if it was doing something, or not.

Yes I agree on 9000 series is a too condensed design causing heat by proximity.

I said that since early days simply because I tear up many cards in the past and seen the new 9000 PCBs on TPU, was easy to make a comparison.



Maybe the AIB want to cover with a smaller cold plate GPU core and VRAM, cheaping out on the metal. The ideal should have been a different VRAM plate being contacting only heatsink and not the GPU core cold plate. But we can't fix the PCB design. And yes IMO they "botched" the design.

Nvidia 5000 series have similar issues but, GDDR7 runs on lower Watt and has lower thermal resistance, therefore easier to cool but, as you know VRAM temps on 5070 Ti aren't great either, better than on 9070XT though.
Also as you can see in the picture VRAM on 5070 is a bit further away while on 9000 is closer to GPU core but, also we have VRM blocks closer to VRAM. VRM doesn't run cool on 300W so again, heat by proximity.



So we have in our grasp this:

1. modify fan curve
2. cool you card from the other side (backplate) - fan in pull with or without heatsink or copper plate. Targeted airflow beats the case fans, don't need to agree but try it.
3. UV
Or combine of the above.

 

[Krit]

HEVC still has better detalization than AV1 at the same bitrate. No reason to choose AV1 at least for Radeon 7000/9000 Series. I would say HEVC is at least ~10-15% better in terms of what you can see on screen. Also for screen recording OBS is way better than Radeon Software it will eat close to 5x less fps while recording.

AV1 10000kbps @2560x1080

HEVC 10000kbps @2560x1080

 

[droopyRO]

Yes I agree on 9000 series is a too condensed design causing heat by proximity.

AMD RX 9070 XT Update: Transition to Samsung GDDR6 Memory for Improved Thermal Performance

The AMD Radeon RX 9070 XT has recently undergone an update involving a change in the type of memory used. Historically, the model has faced criticism for high temperatures associated with its GDDR6 memory, particularly in configurations that adhere closely to the manufacturer's suggested retail...

www.guru3d.com

The transition to Samsung memory has yielded noticeable thermal advantages. Early reviews and comparisons indicate that the new memory modules help lower temperatures significantly. Custom RX 9070 XT models running demanding benchmarks like Furmark have demonstrated memory temperatures of 75°C or lower without necessitating an increase in voltage or a reduction in clock speeds. Notable manufacturers, including Sapphire, XFX, and GIGABYTE, have already integrated Samsung's GDDR6 into their custom variants. However, as of now, AMD has not authorized its partners to explicitly differentiate between models with SK Hynix and Samsung memory, likely to avoid confusion among consumers and maintain a consistent product lineup.

Those Hynix chips are running hotter too.

 

[AusWolf]

AMD RX 9070 XT Update: Transition to Samsung GDDR6 Memory for Improved Thermal Performance

The AMD Radeon RX 9070 XT has recently undergone an update involving a change in the type of memory used. Historically, the model has faced criticism for high temperatures associated with its GDDR6 memory, particularly in configurations that adhere closely to the manufacturer's suggested retail...

www.guru3d.com

Those Hynix chips are running hotter too.

The question isn't whether Hynix chips run hot or not - we've already established that they do. The question is whether it's actually a problem or not.

 

[Scattergrunt]

1. modify fan curve
2. cool you card from the other side (backplate) - fan in pull with or without heatsink or copper plate. Targeted airflow beats the case fans, don't need to agree but try it.
3. UV
Or combine of the above.

I think a lot of more knowledgeable RX 9000 owners are already UVing their cards for the gains it gives. RX 9000 UV's like a monster as much as it OC's, so you can tweak around with the card to your hearts content (being hyperbolic). Doesn't necessarily mean it will fix the issue if you want more performance though, but I think that's why we haven't seen nearly as many people doing it.

 

[Krit]

The question is whether it's actually a problem or not.

I personally wold choose 90C VRAM and 1150rpm fan speed over 80C VRAM and 2000rpm fan speed any day of the week. If it ain't broken then don't fix it.

 

[AusWolf]

I personally wold choose 90C VRAM and 1150rpm fan speed over 80C VRAM and 2000rpm fan speed any day of the week. If it ain't broken then don't fix it.

Exactly. I still stand by my original point - I can't see much chance that all AIBs messed up all of their cards. It can happen with one. Maybe two. But all of them? Nah, c'mon...

 

[Event Horizon]

HEVC still has better detalization than AV1 at the same bitrate. No reason to choose AV1 at least for Radeon 7000/9000 Series. I would say HEVC is at least ~10-15% better in terms of what you can see on screen. Also for screen recording OBS is way better than Radeon Software it will eat close to 5x less fps while recording.

AV1 10000kbps @2560x1080
View attachment 396357
HEVC 10000kbps @2560x1080
View attachment 396358

That's interesting. I've been using HEVC for Moonlight but only because the AV1 support is "experimental". I'll try and compare the image quality when I get a chance. Mind you my bitrate will likely be higher.

 

[Kwadratowicz]

I personally wold choose 90C VRAM and 1150rpm fan speed over 80C VRAM and 2000rpm fan speed any day of the week. If it ain't broken then don't fix it.

For 3 months I will have throttled down GPU cause of CPU, so for now I can make some fun with that card outside the specs .
and even with power limit set to 10, 15, 30 it still uses in benchmark 54W.

Exactly. I still stand by my original point - I can't see much chance that all AIBs messed up all of their cards. It can happen with one. Maybe two. But all of them? Nah, c'mon...

It's not about Them screwing up I'm for example hang on the past, where everything in My PC for the last 6 years had temperatures below 70-75C, and with that temperatures it was silent, now that new hardware is that hot I don't like it and it is harder to cool it like lower power hungry old hardware.

 

[Krit]

That's interesting. I've been using HEVC for Moonlight but only because the AV1 support is "experimental". I'll try and compare the image quality when I get a chance. Mind you my bitrate will likely be higher.

Biggest difference is when in motion static image will show nothing!

 

[biffzinker]

The power limit for this Gigabyte 9070 OC is 245W as stock over the 220W. Also FTM enabled gets me more bandwidth than a manual Mem overclock to 2,800. With FTM enabled at stock 2,518 MHz it scored 923.18 GB/s versus 880.23 GB/s for the 2,800 MHz FTM disabled. Wonder if it can be stable with the 2,800 and FTM enabled. Also trying out a -40Mv on the GPU.

Memory Reads, and Writes is the slot bandwidth so PCIe Gen 5 is enabled on the x16 slot.

 

[NSR]

Roughly a month ago I made this thread talking about crashing issues in Cyberpunk. It came down to most-likely being a PSU problem and Cyberpunk causing high transient spikes. I am finally now going to purchase a new PSU. On that thread a few people have recommended a Seasonic ATX 3.1 PSU. Any other recommendations or is this the way to go?

Is anyone else doing a 3,322/26 MHz boost clock on their 9070? It seems the GPU I have can take the bump from 2,700 to 3,322. Another 10% to the PL shows 269.5W

I did a short stress test and it passed.

Are you talking about the GPU core clocks? I don't think I've gone as high as 3,300 but yes, it easily bumps into the 3000s when undervolted.

 

[Kwadratowicz]

@NSR You can read a little on the power supply tier list on https://docs.google.com/spreadsheet...RB6Xt6JWkc/edit?gid=1719706335#gid=1719706335
I've got for My 9070 xt https://hwbusters.com/psus/super-flower-leadex-vii-gold-1000w-sf-1000f14xg/11/ Super flower PSU. (It works without any problems )
edit1 hwbusters have a lot reviews of psu

 

[AusWolf]

It's not about Them screwing up I'm for example hang on the past, where everything in My PC for the last 6 years had temperatures below 70-75C, and with that temperatures it was silent, now that new hardware is that hot I don't like it and it is harder to cool it like lower power hungry old hardware.

Just because you don't like it because you live in the past, it doesn't mean it's wrong.

Edit: "I don't like the idea of sitting in blazing fast metal boxes that work with explosions. Horses are much safer." See what I mean?

 

[iSpeakVeryWell]

Just finished a quick test using FurMark and HWiNFO. I started with -30% PL. Waited until temps stabilized. To know they have stabilized, I used the reset button of HWiNFO. Waited 2 minutes after a reset, if current temps == average temps, temps are stabilized and I raised PL a notch.

During this quick test, UV was -50mV and VRAM frequency was 2750MHz using fast timing. GPU frequency offset was 0 and I had a linear fan curve.

The card : AsRock 9070 XT Taichi

-30% PL -> 74°C vram / 61°C HS / TBP 238W
-20% PL -> 76°C / 66°C / 272W
-10% PL -> 78°C / 71°C / 306W
0% PL -> 82°C / 76°C / 340W
+10% PL -> 84°C / 82°C / 374W

I believe the jump of 4°C from -10% to 0% is due to lack of precision of the sensor. For example 78.8 could be shown as 78°C. 74.2 -> 76.6 -> 78.8 -> 82 -> 84.2 (could be something else too)

Looks like every time GPU Hot Spot rose 5°C, vram temps rose 2.2°C

There is nothing scientific in that test, it was for fun only.

 

[AusWolf]

Just finished a quick test using FurMark and HWiNFO. I started with -30% PL. Waited until temps stabilized. To know they have stabilized, I used the reset button of HWiNFO. Waited 2 minutes after a reset, if current temps == average temps, temps are stabilized and I raised PL a notch.

During this quick test, UV was -50mV and VRAM frequency was 2750MHz using fast timing. GPU frequency offset was 0 and I had a linear fan curve.

The card : AsRock 9070 XT Taichi

-30% PL -> 74°C vram / 61°C HS / TBP 238W
-20% PL -> 76°C / 66°C / 272W
-10% PL -> 78°C / 71°C / 306W
0% PL -> 82°C / 76°C / 340W
+10% PL -> 84°C / 82°C / 374W

I believe the jump of 4°C from -10% to 0% is due to lack of precision of the sensor. For example 78.8 could be shown as 78°C. 74.2 -> 76.6 -> 78.8 -> 82 -> 84.2 (could be something else too)

Looks like every time GPU Hot Spot rose 5°C, vram temps rose 2.2°C

There is nothing scientific in that test, it was for fun only.

Nice test.

Did you see any change in VRAM clock?

One reason for the temperature change could also be that at lower frequencies, the GPU is asking for less work from the VRAM, too.

 

[iSpeakVeryWell]

With FTM enabled at stock 2,518 MHz it scored 923.18 GB/s versus 880.23 GB/s for the 2,800 MHz FTM disabled.

You probably are triggering ECC, which would dramatically degrade bandwidth. ECC will prevent crashes, but will do it at the cost of performance.

Did you see any change in VRAM clock?

It stayed the same all along. Furmark is still runing and when I go from +10% to -30%, memory controler load go from 24% to 21%.

 

[lexluthermiester]

You probably are triggering ECC, which would dramatically degrade bandwidth. ECC will prevent crashes, but will do it at the cost of performance.

There would have to be a metric ton of errors going on for that much bandwidth degradation to happen.

 

[AusWolf]

It stayed the same all along. Furmark is still runing and when I go from +10% to -30%, memory controler load go from 24% to 21%.

Then it's probably heat soak from the GPU like you said, or less interaction between the GPU and VRAM like I said, or a combination of both.

 

[Fluffmeister]

AMD RX 9070 XT Update: Transition to Samsung GDDR6 Memory for Improved Thermal Performance

The AMD Radeon RX 9070 XT has recently undergone an update involving a change in the type of memory used. Historically, the model has faced criticism for high temperatures associated with its GDDR6 memory, particularly in configurations that adhere closely to the manufacturer's suggested retail...

www.guru3d.com

Those Hynix chips are running hotter too.

Good find, nice to see confirmation at least that these Samsung readings in GPU-Z aren't some misread and they also confirm user findings that they do indeed run a fair bit cooler without needing any tweaks at all. I wonder if in say 6 months or so all 9070/XT people buy will use Samsung VRAM.

The question isn't whether Hynix chips run hot or not - we've already established that they do. The question is whether it's actually a problem or not.

Yeah I doubt it's an issue at all, they run hotter but still well within spec as W1zz confirmed.

I personally wold choose 90C VRAM and 1150rpm fan speed over 80C VRAM and 2000rpm fan speed any day of the week. If it ain't broken then don't fix it.

Yeah can't argue with that, though I can't lie... I kinda like having 80C VRAM at 1150rpm. Though that was clearly luck over judgment.