Friday, April 18th 2025

Chinese Firms Equip GeForce RTX 5090D with Blower-Type Coolers for AI Workstations
Chinese hardware enthusiasts and AI researchers are rolling out custom versions of NVIDIA's latest flagship GPU, the GeForce RTX 5090D "Blackwell" variant for China. Originally intended as a special model that complies with US export regulations, the RTX 5090D is proving to be an excellent choice for smaller AI labs. Building on the trend set by specialist firms modifying the RTX 3090 and RTX 4090, these teams have taken the GB202 chip and mounted it on new PCBs with blower‑style coolers better suited to AI server racks than gaming rigs. Instead of large fans and complex heatsinks, each card uses a dual‑slot blower design that pushes hot air straight out the back of the chassis.
The power connector has also been moved to the rear bracket, making it easier to stack cards in clusters without cables getting in the way. Each unit still packs 32 GB of GDDR7 memory and a PCIe 5.0 x16 interface, although the maximum power draw remains unconfirmed. While previous "D" series cards topped out around 450 W, with Blackwell, this figure might approach 575 W, similar to the regular RTX 5090. This blower cooler could be undersized for a near‑600 W TDP, but with undervolting and underclocking, it may operate with stability over long workloads. Given that Chinese AI labs are struggling to acquire enough GPU capacity, some smaller labs and university researchers are possibly far away from getting enough compute resources for their work. So, repackaging these GPUs into workstation bodies is the only choice to get some acceleration in the hands of the masses.
Sources:
BiliBili, VideoCardz
The power connector has also been moved to the rear bracket, making it easier to stack cards in clusters without cables getting in the way. Each unit still packs 32 GB of GDDR7 memory and a PCIe 5.0 x16 interface, although the maximum power draw remains unconfirmed. While previous "D" series cards topped out around 450 W, with Blackwell, this figure might approach 575 W, similar to the regular RTX 5090. This blower cooler could be undersized for a near‑600 W TDP, but with undervolting and underclocking, it may operate with stability over long workloads. Given that Chinese AI labs are struggling to acquire enough GPU capacity, some smaller labs and university researchers are possibly far away from getting enough compute resources for their work. So, repackaging these GPUs into workstation bodies is the only choice to get some acceleration in the hands of the masses.
41 Comments on Chinese Firms Equip GeForce RTX 5090D with Blower-Type Coolers for AI Workstations
Now this is meant for at AI workstation where noise might not be a big problem. But i don´t think i wanted that thing in my own pc.
I ... have a remedy for that :p
I haven't seen the heatsink but with this fan at 7K RPM, it's technically doable without making the room cold. Just not comfortable for most people.
//The number 230 W is 1080 Ti's die-only power at a serious OC, 471 sqmm is also 1080 Ti's die size. I tested this exact model under these exact conditions in a room which was roughly 60 to 65 °F. 6500 or so RPM was barely enough for it to stay sane. 7K kept it under the boiling point indefinitely. Loud as heck but it worked.
//I would love to have a 4080 with this exact cooling. Should be easy to do it at under 4K RPM.
Still, I enjoy the idea of aftermarket cooler setups being a thing beyond the typical waterblock fare. Imagine building out a GPU by buying a reference board and pairing it up with a custom fin stack, shroud, fans... it's a fun concept, if impractical.
As an example, I run my 2x3090s at 275W each, instead of their default of 370W. The best efficiency for a 3090 is around the 250~300W mark, depending on your task:
benchmarks.andromeda.computer/videos/3090-power-limit
For the 4090 it was around the 300W mark (don't have any easily available link for that one, sorry).
The 5090 seems to have a sweet spot at 400~450W:
LocalLLaMA/comments/1jr6wu2
You did not just abuse affinity law #2, did you? :wtf:
Seriously, increasing the fan's diameter alone would not have such a huge effect on heat dissipation. It does to airflow and pressure, yes, but those don't linearly affect dissipation either.
These are the same people that effectively put a floor on the price of used RTX 4090 ever since the RTX 4090 China export ban. I am pretty sure they played a part in keeping the price of new RTX 4090 as high as it was for the last 18 months by ensuring that demand always exceed supply. There are buyers in Singapore and Malaysia that are actively buying up used RTX 4090, it doesn't matter the model or condition, as long as it is working they are willing to pay around USD2,500 in cash for each card no matter the quantity.
They don't just stop at putting blower‑style cooler onto the modified cards, they are also putting 48GB of VRAM onto the modified PCB to create custom RTX 4090 cards. Nothing is wasted in the process, the discarded parts of the original graphics card are then put back together in their original retail packaging and sold as for display only dummy units. Unfortunately these "corpse cards", missing GPU and VRAM, are starting to show up in places outside of China and usually involved in cases of fraud. I actually just bought a white edition Zotac RTX 4090 AMP Extreme Airo corpse card for USD30 shipped, so I can transfer over the white bits and have a white card.
For whatever reason, they are also buying up used RTX 3090. Although they are not paying as much for them, but this still help create a demand for used RTX 3090. Partly explain why I was able to sell my used RTX 3090 for USD800 in March, and in doing so made a 200 bucks profit over what I paid for it in 2022. It sure doesn't make sense to me when look through the lens of a normal consumer that is using a graphics card to play games, but apparently there is a large market and demand outside of gaming for these cards. By the looks of things, it is a big enough business because the amount of money involved is sizable.
Feels a touch like outrage culture to have a go at these as if enthusiast gamers want them in their study.
I have never seen a thermal issue, melted power connector, or any such nonsense that PC gamers keep encountering. It just doesn't happen. You also have to understand the lift. The laptop you are remoting into things with isn't that strong. The 4 GPU workstation under your desk kinda is. That room full of GPUs is the big fish in all this. There's easily hundreds to thousands of GPUs in a temperature controlled area. I typically wear a track jacket in the office as it's cold. Anyone working in server rooms is probably wearing hearing protection anyways as those 80mm screamers inside the networking gear are by far the worst offenders out there.
Actual professional computing has virtually nothing in common with PC gaming. PC gamers are pretty much always wrong about what high end computing actually is and build crappy systems and screw them up. There's even a hiring bias now about this. We game at work but on PS5 and Switch. If you mention PC gaming in an interview and it's not Doom II or Quake era stuff you're not getting hired or if you do it's a five figure less pay rate per yer as you've outed yourself as a moron.