NVIDIA GeForce RTX 5090D Overclocked to a Staggering 3.4 GHz and 34 Gbps Memory

[AleksandarK]

Yes, the title is correct. One of NVIDIA's GeForce RTX 5090D "China" edition GPUs, not the regular RTX 5090, managed to run at 3.4 GHz under liquid nitrogen. With a staggering 575 W default TDP, Tony Yu, ASUS China's general manager, has performed physical modifications that allow the card to run up to 1000 W TDP. The RTX 5090D is a China-exclusive variant with virtually no difference from the regular RTX 5090, just limited general AI capability due to US export regulations. ASUS China used its top-end Astral OC variant for this stunt, which, as we proved in our review of the regular ASUS RTX 5090 Astral OC, has some pretty good chip binning, allowing the card to reach the highest overclock. We pushed the regular RTX 5090 Astral OC GPU on air to 3086 MHz, a +277 MHz over the stock boost setting. However, the RTX 5090D equivalent under LN2 manages to reach 3,390 MHz at peak loads, which is a +581 MHz difference.

For memory, the overclock is equally impressive with 34 Gbps. Regarding performance, the LN2-overclocked RTX 5090D surpassed stock performance by approximately 16%. During benchmark tests, the GPU outperformed multiple previous-generation graphics cards, including a dual RTX 3090 Ti configuration in Port Royal and a quad GTX 1080 Ti setup in Fire Strike. Power consumption figures indicate that 1,760 W was used in total for a rig with ASUS ROG Astral RTX 5090D, which is paired with the Ryzen 7 9800X3D on the ASUS ROG X870E Hero motherboard. This roughly yields a 1,000 W power consumption by the card, which has seen its PCB get physical modifications to output such high power.

View at TechPowerUp Main Site | Source

 

[redeye]

Any one going to joke… “ASUS (china) got the power”
seriously… 34Gbps? , why not install the 36gbps stock versions of RAM. If they can swap ssd chips in a iPhone for larger memory, certainly can change the RAM…

or more importantly can I get a 5090d at a discount? Is that not what the “d” stands for?

 

[TumbleGeorge]

allowing the card to reach the highest overclock. We pushed the regular RTX 5090 Astral OC GPU on air to 3086 MHz, a +277 MHz over the stock boost setting.

Stock boost of this model is 2580... Difference to 3086 isn't 277.

 

[N/A]

Stock boost of this model is 2580... Difference to 3086 isn't 277.

actually 2809 boost is the average for 25 games, peak is 2932.

 

[freeagent]

Is stock boost even a thing? Mine says it is supposed to boost to 2625 but really its like 2950.

 

[W1zzard]

Yeah, I got 34 Gbps too, on all my cards that I reviewed .. at that point the NVIDIA memory overclock limiter kicks in and won't let you go further, even though it's stable.

But that confirms that ASUS has no control over that limiter, even when they are able to adjust the VBIOS power limit beyond 600 W and are still able to sign the BIOS

 

[TumbleGeorge]

actually 2809 boost is the average for 25 games, peak is 2932.

There is an ambiguity here. What is called "stock boost" in the article:

the stock boost setting.

and what is achieved in games, as described in the review, are different things.

 

[Lew Zealand]

There is an ambiguity here. What is called "stock boost" in the article:

and what is achieved in games, as described in the review, are different things.

It's been that way since Pascal, so 8+ years. Nvidia has a listed boost but every card boosts quite nicely above that number.

 

[TumbleGeorge]

Nvidia has a listed boost

Yes 2407MHz. 2580MHs is boost number from Asus in specifications of the model.

 

[Lew Zealand]

Yes 2407MHz. 2580MHs is boost number from Asus in specifications of the model.

Read W1zz's review for that (or any) model and see where the core clock typically sits when gaming, it's listed as the Median. It'll be above either of those numbers for any Pascal or later GPU.

Here ya go:

Typical gaming clocks are 2857 MHz for the ASUS GeForce RTX 5090 Astral OC W1zz reviewed.

 

[agent_x007]

Stock boost of this model is 2580... Difference to 3086 isn't 277.

Single "locked" frequency of the core isn't a thing since Kepler.
Since that time, at least on NV, you just get "down to" values for bases that card can throttle to and be marked as "fine" :
1) "Base" frequency : Card must work reliably at this frequency regardless of cooling or limits used under 3D load (if it's not - RMA).
2) "Boost" frequency : "Healthy" card will increase it's core clock to this frequency at the minimum [manufacturer guaranteed] during 3D load - everything beyond that doesn't have to work (but cards will boost to moon since... they can - only voltage limit/power limit/temp limit are the real show stoppers). If card doesn't boost to "boost" frequency provided, RMA card (or try to fix it if it's after warranty).

 

[TumbleGeorge]

Read W1zz's review for that (or any) model and see where the core clock typically sits when gaming, it's listed as the Median. It'll be above either of those numbers for any Pascal or later GPU.

Here ya go:

Typical gaming clocks are 2857 MHz for the ASUS GeForce RTX 5090 Astral OC W1zz reviewed.

Apparently there is a misunderstanding from some colleagues of what I wrote after reading the article, where two different things are mentioned. Namely the official increase in specifications that Asus described and the actual values achieved in the review by Wiz. I have already described several times what I mean... Maybe the problem is that none of us, who cannot understand each other, didn't speak English perfectly. Include me.

 

[sudothelinuxwizard]

or more importantly can I get a 5090d at a discount? Is that not what the “d” stands for?

D stands for Dragon IIRC.

 

[lexluthermiester]

500mhz out of 2700mhz is staggering? Seriously?

Yeah, I got 34 Gbps too, on all my cards that I reviewed ..

And then there's this.

I mean, it's a good effort, but it's not very remarkable, certainly not staggering.

 

[Bwaze]

“NVIDIA GeForce RTX 5090D Overclocked to a Staggering 3.4 GHz…
With a staggering 575 W default TDP…”

stagger
verb
1. walk or move unsteadily, as if about to fall.

 

[lexluthermiester]

“NVIDIA GeForce RTX 5090D Overclocked to a Staggering 3.4 GHz…
With a staggering 575 W default TDP…”

stagger
verb
1. walk or move unsteadily, as if about to fall.

Ah! Well in that context, maybe.

 

[N/A]

GDDR7 can probably do 2.5 GHz 40 Gbps already and Nvidia is hiding it.

 

[x4it3n]

I mean, it's a good effort, but it's not very remarkable, certainly not ststaggering.

Apparently it's the max the slider can go to, it's locked (as of now at least). But 21% is already a great improvement over stock! That's almost 2.2TB/s of Bandwidth!

 

[N3utro]

 

[x4it3n]

GDDR7 can probably do 2.5 GHz 40 Gbps already and Nvidia is hiding it.

GDDR7 can go up to 40Gbps yeah but those are probably binned chips. The max at Stock is 32Gbps right now, and even at 32Gbps the yields were surprisingly not good enough, hence the 5080 only having 30Gbps chips. But in all the OC tests I've seen, all the GDDR7 chips have been able to reach 34Gbps! So I really wonder why no 32Gbps chips in RTX 50s. Maybe they degrade over time? Stange...

 

[tpuuser256]

fuu.. we have to do mods to the mcb for any more power. It's frustrating since this clards overclocks relatively well.

 

[x4it3n]

fuu.. we have to do mods to the mcb for any more power. It's frustrating since this clards overclocks relatively well.

Yeah that's why I was surprised none of the AIBs included 2x 16-pin connectors with some 800W BIOS or something! Also that would have also made each connector pull less watts since it could have split the charge to 400W each! But I feel like NVIDIA are the ones who prevented AIBs to do that so more people buy the FE (2-slot/SFF GPU) and can make even more money.

 

[TumbleGeorge]

Yeah that's why I was surprised none of the AIBs included 2x 16-pin connectors with some 800W BIOS or something! Also that would have also made each connector pull less watts since it could have split the charge to 400W each! But I feel like NVIDIA are the ones who prevented AIBs to do that so more people buy the FE (2-slot/SFF GPU) and can make even more money.

In this line of thought, the question arises, how many users have PC cases from which such heat outputs can be derived? It is close to the mind that there may even be enthusiasts with PC cases in which placing a high-power video card can literally bake the other components and itself and the user himself, by blowing all this heat towards his body.

 

[freeagent]

how many users have PC cases from which such heat outputs can be derived?

I have 2 that can for sure, and maybe the other one, but would have to use strong fans.

 

[Papusan]

Yeah that's why I was surprised none of the AIBs included 2x 16-pin connectors with some 800W BIOS or something! Also that would have also made each connector pull less watts since it could have split the charge to 400W each! But I feel like NVIDIA are the ones who prevented AIBs to do that so more people buy the FE (2-slot/SFF GPU) and can make even more money.

That will be for the Titan (5090Ti) cards if it ever come.