Manli Rep. Confirms Downgrading of GeForce RTX 5090D Graphics Cards - Only 24 GB of GDDR7 VRAM

[T0@st]

Around late April, Chinese industry insiders started whispering about a possible halting of NVIDIA GB202 "Blackwell" GPU shipments into the region. Team Green's local board partners released custom "market exclusive" GeForce RTX 5090D 32 GB cards; featuring slightly downgraded flagship silicon. Since launch, Chinese hardcore gaming enthusiasts and DIY AI firms have observed impressive performance credentials, despite the presence of factory limited specifications. On May 6, reports suggested a complete halt of GeForce RTX 5090D sales in China. Fresher rumors indicate NVIDIA's engineering team return to the drawing board; with their alleged preparation of an even weaker GeForce RTX 5090D design.

A group of Baidu tipsters have discussed a key area of (further) compromise: VRAM capacity. The original GeForce RTX 5090D configuration was armed with 32 GB of GDDR7 VRAM; mid-May speculation envisioned a 24 GB variant. Earlier today, harukaze5719 highlighted an intriguing Weibo post. A Manli webshop representative has "confirmed" that revised GeForce RTX 5090D cards will become available around July. Up until recent events, this Chinese brand was selling flagship GPU-based Gallardo (black) and Stellar (white) models. A web chat session was captured and shared on Weibo—in addition, the sales agent disclosed their belief that (NVIDIA's) supply of GeForce RTX 5090D GPUs has been "insufficient since launch" time. On a semi-positive note, customers will not be greeted by price hikes. Manli anticipates "unchanged" price tags; albeit associated with lesser specifications.



View at TechPowerUp Main Site | Source

 

[Hyderz]

this will drive up the price of the current 5090D when the lesser model will sell at the same price when it becomes available

 

[redeye]

Wow. Nvidia “repurposed” GPU’s…

 

[igormp]

I wonder if those will be just a GB203 with 3GB GDDR7 modules - basically the same as the rumored 5080 Super -, or if it'll be a really cut down GB202 with a 384-bit bus. If the latter, then it could also become a future 5080ti.

 

[N/A]

Are they back to releasing vastly different spec under the same naming 5090D with 14080 shader 384 bit. 4080 confusion again.

 

[Bjorn_Of_Iceland]

This is probably what the actual RTX 5080 specs should have been

 

[Dr. Dro]

This is probably what the actual RTX 5080 specs should have been

Nothing of the sort. The 5080 was always intended to use the GB203 chip from the get go, and there is actually a good reason for it, after all. This is intended to meet the U.S. government restrictions, which were made more restrictive after the launch of the 5090D - to be more specific, made official a few days ago.

https://www.cnbc.com/2025/05/14/ai-chip-export-rules-nvidia.html

Beyond the fact that it's practically twice as fast as the RX 9070 XT at rendering games (and that is where most people stop seeing the big picture), for compute, that goes a step further still. The RTX 5090 is extremely powerful. So powerful, in fact, that it is directly subject to export controls. The RTX 5090D model sold in China had its inferencing performance intentionally reduced by around a third, using firmware and chip level tricks, but it is still deemed too powerful for export for restricted group countries, such as China according to the export restrictions currently in effect. Additionally, countries which aren't on the United States' most privileged export tier, but are allowed to have this product (such as Brazil), face a limitation on how many units can be distributed to them. Nvidia can only ship a certain amount of 5090s to them and must obtain authorization to send more should this limit ever be reached.

As stated in the OP, the supply of 5090D was never too good, and it doesn't surprise me to see its specification being revised down. There is nothing Nvidia can do about this, really, except comply and reduce their product's performance until it meets the specifications.

I wonder if those will be just a GB203 with 3GB GDDR7 modules - basically the same as the rumored 5080 Super -, or if it'll be a really cut down GB202 with a 384-bit bus. If the latter, then it could also become a future 5080ti.

Given the name isn't changing to 5080D or something, it's probably going to be the same 5090D core we currently have with the bandwidth and capacity reduced, with perhaps a further worsened ratio for AI TOPS.

 

[SOAREVERSOR]

Nothing of the sort. The 5080 was always intended to use the GB203 chip from the get go, and there is actually a good reason for it, after all. This is intended to meet the U.S. government restrictions, which were made more restrictive after the launch of the 5090D - to be more specific, made official a few days ago.

https://www.cnbc.com/2025/05/14/ai-chip-export-rules-nvidia.html

Beyond the fact that it's practically twice as fast as the RX 9070 XT at rendering games (and that is where most people stop seeing the big picture), for compute, that goes a step further still. The RTX 5090 is extremely powerful. So powerful, in fact, that it is directly subject to export controls. The RTX 5090D model sold in China had its inferencing performance intentionally reduced by around a third, using firmware and chip level tricks, but it is still deemed too powerful for export for restricted group countries, such as China according to the export restrictions currently in effect. Additionally, countries which aren't on the United States' most privileged export tier, but are allowed to have this product (such as Brazil), face a limitation on how many units can be distributed to them. Nvidia can only ship a certain amount of 5090s to them and must obtain authorization to send more should this limit ever be reached.

As stated in the OP, the supply of 5090D was never too good, and it doesn't surprise me to see its specification being revised down. There is nothing Nvidia can do about this, really, except comply and reduce their product's performance until it meets the specifications.



Given the name isn't changing to 5080D or something, it's probably going to be the same 5090D core we currently have with the bandwidth and capacity reduced, with perhaps a further worsened ratio for AI TOPS.

PC gamers are some of the stupidest and most self centered idiots on the planet. Every GPU sold 8800 on upward was a compute product and not a gaming product. That's the reality. PC gaming is going to rent a GPU in the cloud and every PC gamers who's bought on Steam or bought an 8800 or later card voted for that and paid for it to happen. Congrats! You got what you paid and voted for!

 

[Dr. Dro]

PC gamers are some of the stupidest and most self centered idiots on the planet. Every GPU sold 8800 on upward was a compute product and not a gaming product. That's the reality. PC gaming is going to rent a GPU in the cloud and every PC gamers who's bought on Steam or bought an 8800 or later card voted for that and paid for it to happen. Congrats! You got what you paid and voted for!

Well, it wasn't a huge problem until the first crypto bubble happened. In a sense, I think the AI grift is a continuation of that bubble, it's a lot of highly marketable hype with lots of room to make some serious money, so the small guy (in this case, the gamer and content creator) ends up paying the ultimate price.

The only way to dissociate from this problem going forward is to intentionally split consumer and enterprise products across different nodes. The AI market is hungry for high performance accelerators only possible on the latest generation process, and they are willing to pay for this.

 

[LabRat 891]

Something tells me these will make it stateside under another name.

Guessing, RTX 5080 Ti SUPER?

 

[_roman_]

I assume Nvidia put 8GB VRAM on graphic card to segment a useful card from the cheap entry cards.

PC gamers are some of the stupidest and most self centered idiots on the planet.

I'm innocent. I paid a few time short subscriptions to get amazon prime for the games. I never spend any cash on Steam, ubisoft, EA, .... or RGB or fish tank cases or vertical gpu mounts