Sapphire Radeon RX 9070 XT Pulse

[LittleBro]

Are you talking about the very very gradual decline in clocks from 2800 to 2779 after the initial spike, or the initial spike?
That brief spike to 2910MHz isn't representative of the sustained load, it's not a valid part of the sustained load test, but a necessary part as the test obviously starts from "cold".

To my understanding, power limit should not allow for any excessive power draws, except for power spikes.
This hike to 2910 lasted about 3 seconds, so definitely not a 20ms power spike.

Yes, was talking about gradually lowering clocks while temps were at their highest.

This is what I would expect from fan speed behavior (look at the graph at the bottom):

Radeon RX 9070 XT (ASUS TUF OC) review (Page 6)

We review the Radeon RX 9070 XT (ASUS TUF OC). The most striking improvement lies in the number of AI and ray tracing (RT) accelerators, which have seen a noticeable boost compared to previous generations. The RX 9070 XT packs 64 RDNA 4 compute units, 64 RT accelerators, and 128 AI accelerators.

www.guru3d.com

They also mention throttling. Damn pity we don't have all 3 temps with load, power draw and clocks in same graph.

 

[W1zzard]

Yes, was talking about gradually lowering clocks while temps were at their highest.

This is normal, when the GPU is cold it will clock itself much higher and then gradually lower clocks as it heats up.

That's also why you should let the card heat up for a bit once you've loaded into your benchmark, for every single run, before taking any data. Most reviewers start taking data immediately, so their results include those higher clocks, which you get for a few seconds when actually playing the game and then never again.

This is also a problem with integrated benchmarks, because they have no warmup phase

This is what I would expect from fan speed behavior (look at the graph at the bottom):

That's just warm card plotted over time, not from a cold card, so clocks won't move. It's basically the red part of the chart below, i.e. a single-snapshot of relevant data, no need for an actual chart

http://imgur.com/6P0ArfK

 

[LittleBro]

This is normal, when the GPU is cold it will clock itself much higher and then gradually lower clocks as it heats up.

That's also why you should let the card heat up for a bit once you've loaded into your benchmark, for every single run, before taking any data. Most reviewers start taking data immediately, so their results include those higher clocks, which you get for a few seconds when actually playing the game and then never again.

This is also a problem with integrated benchmarks, because they have no warmup phase

Thanks for explanation. Okay, so clocks ended up 200 MHz below max. boost clocks because of thermal limit, right?
The temps (GPU hotspot, VRAM) are at their highest values and fans are at their lowest. Maybe it's quiet but definitely not cool.
To me, this fan speed behavior may actually become a reason for performance bottleneck in the long run.

Have you seen this fan behavior before? I've owned several Sapphire cards and all of them were incrementing fan speed until GPU temps stopped rising (not talking about hitting temp limit now).

 

[rahlquist]

Wizzard, thank you for doing such thorough reviews with the amazing teardowns and photos. A question, like a few other old timers, I tend to hate seeing fan curves letting some temps, like the ram on these get so high. The only public spec Hynix shares for these that I have found shows a operational temp range of "0~85c" with no mention of what they consider to truly be the TJMax. I've found people reporting memory temps on the 9070 family up to 105c. When I brought up wondering if sapphire had mentioned plans for an improved vbios for the cards over on AMD's discord and several other places I get the wave off saying the chips are "capable of far higher temps than I think."

I was wondering if you had ever run across what the actual TJMax is for these DDR6 chips? I have my fan curves tweaked slightly, just enough to keep the RAM below 90c with furmark running but it still bothers me the blase attitude about this. I know these chips have been around for a while, but it would still be one of those things that are good to know for certain.

Anyhow thanks for all you do!

 

[W1zzard]

Thanks for explanation. Okay, so clocks ended up 200 MHz below max. boost clocks because of thermal limit, right?

Not because of the thermal "limit", just because of warming up, it's not hitting any limit.

Maybe it's quiet but definitely not cool.

Shows 45°C

Have you seen this fan behavior before?

Yeah this happens from time to time, mostly on AMD cards, due to limitations on how their fan curve can be configured. Basically the card heats up an the GPU guesses it will heat up very quickly, so it spins the fans super high, but then realizes it overdid it and slows them down again, but too much, so it goes into this kind of oscillation that you see

what the actual TJMax is for these DDR6 chips?

95°C case 24/7, so tjmax will be like 120+

 

[rahlquist]

Yeah this happens from time to time, mostly on AMD cards, due to limitations on how their fan curve can be configured. Basically the card heats up an the GPU guesses it will heat up very quickly, so it spins the fans super high, but then realizes it overdid it and slows them down again, but too much, so it goes into this kind of oscillation that you see

Oh it clearly needs m0ar AI.

BTW thanks for the Tjmax info!

 

[LittleBro]

Not because of the thermal "limit", just because of warming up, it's not hitting any limit.

If it's not hitting any limit, why doesn't it boost any higher than 2780 MHz during sustained full load?

If I understand the concept of modern GPU clocks funcionality correctly, the GPU will try to maintain maximum boost until:
- max. boost freqeuncy limit is met, or
- max. power draw limit is met, or
- max. temp limit is met, or
- CPU becomes a bottleneck for GPU.

If Pulse is boosting during full load to just 2780 MHz, I guess TBP is already 304W and that's preventing further boosting.

Shows 45°C

I'm talking GPU hotspot and/or VRAM temps here.

I'd do the testing myself, unfortunately my 9070 XT is still on its way to be delivered ... /facepalm

 

[W1zzard]

the GPU will try to maintain maximum boost until

That's not exactly right. Both modern AMD and NVIDIA GPUs will gradually boost less high as temperature and other factors increase, once they hit a certain "limit" temperature they will throttle much harder in addition to that, and once they hit another limit they will shut off/restart/crash

 

[AusWolf]

If it's not hitting any limit, why doesn't it boost any higher than 2780 MHz during sustained full load?

If I understand the concept of modern GPU clocks funcionality correctly, the GPU will try to maintain maximum boost until:
- max. boost freqeuncy limit is met, or
- max. power draw limit is met, or
- max. temp limit is met, or
- CPU becomes a bottleneck for GPU.

If Pulse is boosting during full load to just 2780 MHz, I guess TBP is already 304W and that's preventing further boosting.

The Pulse has a power limit of 315 W, but besides that, different scenes put a different load onto your GPU, so even when hitting a steady 304 or 315 W power wall, you'll see fluctuating clocks. There aren't two frames that stress every single component in your GPU equally. GPUs are complex things, and modern graphics are complex to render, too.

Add what W1zz said about temperatures into the mix, and you'll see why a modern GPU (especially an AMD one) never flatlines on clock speed.

For example, my 9070 XT boosts to 3.1 GHz in some scenarios, 3 GHz in others, but just 2.8 GHz in yet others. There are far too many variables for a regular home user to imagine.

 

[W1zzard]

never flatlines on clock speed

It actually does once you are in-game for a while, temperatures and fan speeds have stabilized, and you don't move around, confirms that the clock selection is not random but deterministic

 

[LittleBro]

The Pulse has a power limit of 315 W, but besides that, different scenes put a different load onto your GPU, so even when hitting a steady 304 or 315 W power wall, you'll see fluctuating clocks.

As for fluctuating clocks, they should not fluctuate +/- 100 MHz when load is steady. I'm not talking about 100% load but a same constant load, like W1zzard mentioned.
In this type of load, clocks would fluctuate by a huge margin only due to possible throttling.

For example, my 9070 XT boosts to 3.1 GHz in some scenarios, 3 GHz in others, but just 2.8 GHz in yet others. There are far too many variables for a regular home user to imagine.

I'm fully aware of a fact that GPU clocks may vary depending on type of instructions (load) being used. This is clear. I'm using same type of load for comparison - W1zzard's type of load.


I'm trying to find what's limiting the GPU to achieve max. boost clock in sustained load tests published in reviews.

Added clocks to the table:

Cards with higher TBP tend to boost higher in W1zzard's tests. There's clearly a correlation. It seems there's a slight relation between clocks and raising temps as well. Logically, these cards must be hitting power limit during the tests. This is backed by a fact that once you undervolt a card, it will immediately boost higher (undervolting creates power headroom). Increasing Power Limit and frequency offset seems to be broken to a some sort of degree, as pointed out in other forum thread - it does not always yield better performance. Which is strange, because performance is relating to clocks, those are in relation with power limit, therefore if you increase PL along with clock limit, performance normally increases and not the opposite.

Also check out XFX's card fan behavior in the graphs below. This is behavior I'd consider standard. You don't lower fans when temp only keeps raising.

Waiting for ASUS card's review to be published.

Spoiler: GPU graphs

 

[Lew Zealand]

I'm trying to find what's limiting the GPU to achieve max. boost clock in sustained load tests published in reviews.

Could be a combination of silicon lottery and/or binning better chips for the non-MSRP models. Also could be simple power draw: Taichi with silent bios only draws a little more power and gets only a little faster. Higher power limits enable higher clocks.

Unfortunately with only a single Pulse data point we can't know whether it's any or all of these.

 

[AusWolf]

As for fluctuating clocks, they should not fluctuate +/- 100 MHz when load is steady. I'm not talking about 100% load but a same constant load, like W1zzard mentioned.
In this type of load, clocks would fluctuate by a huge margin only due to possible throttling.


I'm fully aware of a fact that GPU clocks may vary depending on type of instructions (load) being used. This is clear. I'm using same type of load for comparison - W1zzard's type of load.


I'm trying to find what's limiting the GPU to achieve max. boost clock in sustained load tests published in reviews.
View attachment 392941

Added clocks to the table:
View attachment 392942

Cards with higher TBP tend to boost higher in W1zzard's tests. There's clearly a correlation. It seems there's a slight relation between clocks and raising temps as well. Logically, these cards must be hitting power limit during the tests. This is backed by a fact that once you undervolt a card, it will immediately boost higher (undervolting creates power headroom). Increasing Power Limit and frequency offset seems to be broken to a some sort of degree, as pointed out in other forum thread - it does not always yield better performance. Which is strange, because performance is relating to clocks, those are in relation with power limit, therefore if you increase PL along with clock limit, performance normally increases and not the opposite.

Also check out XFX's card fan behavior in the graphs below. This is behavior I'd consider standard. You don't lower fans when temp only keeps raising.

Waiting for ASUS card's review to be published.

Spoiler: GPU graphs

View attachment 392945
View attachment 392946
View attachment 392947
View attachment 392948
View attachment 392949

As far as I'm aware, all modern AMD cards target max power limit. Achieving it is a combination of the nature of the load, temperature and clock speed, which all fluctuate in a live game. Like W1zz said, the only way to see steady clocks is by rendering a static scene/image with equalised thermals.

The last card I've seen from AMD that hit a max boost clock limit before hitting max power was the 6500 XT.

 

[Ninehell]

Got Sapphire official in my country said they start selling around middle of this month. Now I'm getting my cash ready.

 

[LittleBro]

Got Sapphire official in my country said they start selling around middle of this month. Now I'm getting my cash ready.

What are they gonna start selling? Pure pulse drugs?

 

[Ninehell]

Only see Pulse and Nitro+ but still thinking which one to grab. I like the Nitro+ look but leaning toward Pulse cause 2x8 pins and no RGB.

All other major PC stores here has 9070xt page created on their site, but only answer me with "We don't know", "We don't sell it here", "Check again next month".
Only Sapphire told me when to expect.

 

[LittleBro]

WTH? They already started selling RX 9070s.

 

[Oasis]

Does anyone know what thickness the memory thermal pads this card has? I have one and want to change them out for better temps
2mm?

https://www.techpowerup.com/review/sapphire-radeon-rx-9070-xt-pulse/images/heatpipes.jpg

Thanks

@W1zzard