LN2‑Cooled GALAX GeForce RTX 5090D Reaches 3,650 MHz Core, 36 Gbps GDDR7

[AleksandarK]

Team OGS has demonstrated an impressive feat of extreme overclocking by extracting every last drop of performance from GALAX GeForce RTX 5090D HOF edition. Armed with a custom Extreme OC BIOS, authorized for up to 2,000 W of power, and two 12V-2x6 connectors delivering 1,200 W, the group cooled the card with liquid nitrogen and fine-tuned voltages to push clock speeds far beyond factory limits. In the GPUPI 32B benchmark, the GPU achieved a record-breaking core frequency of 3,650 MHz and a completion time of just 39.434 seconds, securing a new world record for GPU-accelerated Pi calculation.

The team's success extended to graphics tests as well. In 3DMark Port Royal, the GPU operated at 3,570 MHz on its core and 2,250 MHz on its GDDR7 memory, achieving 47,469 points. Under Unigine Superposition's 1080p Extreme preset, they maintained 3,540 MHz and 2,250 MHz, respectively, reaching 38,237 points, an increase of several thousand over previous records. Pushing memory speeds to an effective 36 Gbps translates into 2.304 TB/s of bandwidth, nearly 29% above the standard specifications. These results show the untapped potential of NVIDIA's Blackwell architecture and suggest that, with similar power, cooling, and VBIOS modifications, professional cards like the RTX Pro 6000 series could achieve even more significant gains, given the higher CUDA core count.



View at TechPowerUp Main Site | Source

 

[Tomgang]

Oh i can easily beat that with my superior rtx 4090. I can beat it on stock air cooling even blind folded

Wait why are you all looking at me like that......

 

[igormp]

Pushing memory speeds to an effective 36 Gbps translates into 2.304 TB/s of bandwidth, nearly 29% above the standard specifications.

Wew, that's a lot of bandwidth. Those GDDR7 modules sure can clock high.

 

[Tomorrow]

Wew, that's a lot of bandwidth. Those GDDR7 modules sure can clock high.

Not really. 36Gbps is very low considering every single 5080 can achieve 34Gbps on air with stock coolers and rest can achieve 32Gbps.
That's because Nvidia has locked G7 OC. All models go to one number and that's it.

I wonder what they did on that board to bypass the 32Gbps limit 5090 normally has.
Still with LN2 i would have expected more than a mere 36Gbps.

 

[MentalAcetylide]

Sounds interesting & all that, but when I saw "liquid nitrogen", it lost all relevancy to me.
These results show the untapped potential of NVIDIA's Blackwell architecture and suggest that, with similar power, cooling, and VBIOS modifications, professional cards like the RTX Pro 6000 series could achieve even more significant gains, given the higher CUDA core count.
Untapped potential for...? The only thing I see is the potential for setting new records. I don't think too many of us are going to be sporting LN2-cooled systems to squeeze out that extra performance for shits & giggles in a youtube video, and unless I'm behind the times, I don't think companies are going to be opting for expensive LN2 cooling for that extra performance; they'll just order more cards to make up the performance difference(which would still probably be much cheaper in the long run).
Also, unless they're able to start pushing towards more efficiency, this stuff is just going to keep on getting bigger & more power hungry.

 

[LastDudeALive]

I don't think companies are going to be opting for expensive LN2 cooling for that extra performance

No, but there is obvious potential to make a startup claiming to offer liquid nitrogen cooling for AI servers.

Raise $50 million from VCs who will throw money at anyone saying "AI," acquire the best computers money can buy, hire all your friends, and spend your time playing Halo 3 for the rest of your life.

 

[sarthurn]

Sounds interesting & all that, but when I saw "liquid nitrogen", it lost all relevancy to me.
These results show the untapped potential of NVIDIA's Blackwell architecture and suggest that, with similar power, cooling, and VBIOS modifications, professional cards like the RTX Pro 6000 series could achieve even more significant gains, given the higher CUDA core count.
Untapped potential for...? The only thing I see is the potential for setting new records. I don't think too many of us are going to be sporting LN2-cooled systems to squeeze out that extra performance for shits & giggles in a youtube video, and unless I'm behind the times, I don't think companies are going to be opting for expensive LN2 cooling for that extra performance; they'll just order more cards to make up the performance difference(which would still probably be much cheaper in the long run).
Also, unless they're able to start pushing towards more efficiency, this stuff is just going to keep on getting bigger & more power hungry.

Exactly. Why not just focus on what we could all possibly do. Its bragging how fast your Civis SI is ... on nitro. Who cares. What can you do with stock parts, not some bolt on cheat.

 

[MentalAcetylide]

No, but there is obvious potential to make a startup claiming to offer liquid nitrogen cooling for AI servers.

Raise $50 million from VCs who will throw money at anyone saying "AI," acquire the best computers money can buy, hire all your friends, and spend your time playing Halo 3 for the rest of your life.

Not going to happen. When it comes to commercial cooling, they'll use ammonia long before they would even think about nitrogen. Then you have to keep the humidity down lower than whale shit because moisture of any kind is going to be a problem. Not worth it.

 

[Tomorrow]

I wonder what they did on that board to bypass the 32Gbps limit 5090 normally has.

MSI Afterburner 4.6.6 Beta 6 with unofficial RDNA4 support

As I promised some time ago, today we launch new beta of MSI Afterburner with AMD RDNA4 GPU architecture support. I'd like to explicitly mention one...

forums.guru3d.com

Looks like it's thanks to Afterburner removing Nvidia's limit.

 

[Vayra86]

Sounds interesting & all that, but when I saw "liquid nitrogen", it lost all relevancy to me.
These results show the untapped potential of NVIDIA's Blackwell architecture and suggest that, with similar power, cooling, and VBIOS modifications, professional cards like the RTX Pro 6000 series could achieve even more significant gains, given the higher CUDA core count.
Untapped potential for...? The only thing I see is the potential for setting new records. I don't think too many of us are going to be sporting LN2-cooled systems to squeeze out that extra performance for shits & giggles in a youtube video, and unless I'm behind the times, I don't think companies are going to be opting for expensive LN2 cooling for that extra performance; they'll just order more cards to make up the performance difference(which would still probably be much cheaper in the long run).
Also, unless they're able to start pushing towards more efficiency, this stuff is just going to keep on getting bigger & more power hungry.

Yeah I've always had that with these records, too.

If its not usable 24/7 what's the point. Its a complete waste of time, wee number can go up to 11 look at this I managed to make a screenshot before it actually crashed already. Its not proving a single goddamn thing to anyone anywhere in the history of whatever, all you proved was you have a hobby that you spent time on

The real potential is untapped with actually usable products. Its not like Nvidia is offering a sliver of free performance here like you would from any regular overclock. The whole idea about untapped potential is... lol. This is the GPU that comes in a shroud measuring two slots, limiting its cooling potential by design, but still actually just limited by power envelope and 12v6. So you're going to ghetto mod a 2,5-3,5k worth chip to get to its 'untapped potential' to somehow... profit from it?! The mental asylum is that way

 

[Jism]

I think it's more efficient to directly solder wires onto the 6 pin connection, and be able to push far more then what it was designed for?

The same applies with the PCI-E 8 Pin. Rated at 150W but it can easily chew 450W up to 600W depending on quality of PSU and / or Wires.

I hate that both AMD and Nvidia are locking bios files - you have no other option then to work within it's pre-set limits.

I mean i understand a unlocked card or bios is capable of pushing 1000W's - most VRM's are build with such overhead.

I assume it's all because of saving costs. If you can design a VRM that is lower in regards of capacity and your user is unable to OC past a certain point then you'll need less components to reach the same.

 

[MentalAcetylide]

Yeah I've always had that with these records, too.

If its not usable 24/7 what's the point. Its a complete waste of time, wee number can go up to 11 look at this I managed to make a screenshot before it actually crashed already. Its not proving a single goddamn thing to anyone anywhere in the history of whatever, all you proved was you have a hobby that you spent time on

The real potential is untapped with actually usable products. Its not like Nvidia is offering a sliver of free performance here like you would from any regular overclock. The whole idea about untapped potential is... lol. This is the GPU that comes in a shroud measuring two slots, limiting its cooling potential by design, but still actually just limited by power envelope and 12v6. So you're going to ghetto mod a 2,5-3,5k worth chip to get to its 'untapped potential' to somehow... profit from it?! The mental asylum is that way

Yeah, its an Underpants Gnome-ism. Collect underpants ---> ? ---> Profit! or LN2 cool NVidia graphics card ---> ? ---> Profit!
Now on the other hand, I could understand tweaking stuff in order to get the cards to NOT run at full-blast & save on power consumption when its feasible for a given task, like improving efficiency while sacrificing some performance, although these cards don't really have efficiency in mind, heh.