Brother Zhang fixes AI RTX 5090 card by cutting memory-bus
An interesting card, and an interesting fix.
This RTX 5090 for AI was made at your gaming card expense
This repaired GeForce RTX 5090 is not a normal retail card sent back to service. It appears to be one of the blower-style models assembled in China for AI and server use, using parts taken from standard gaming cards. These products are not made by official board partners in this form. Instead, specialized factories reportedly strip working GeForce RTX 5090 gaming cards, remove the GPU, memory, and supporting components, and transplant them onto a different PCB with a blower cooler.
Designing a proper blower card for every custom RTX 5090 model would take time and money, and each partner board would need its own cooler and validation work. These companies instead appear to rely on one shared board design, then move salvaged components onto that layout. It is a faster and cheaper method, and it allows mass conversion of gaming cards into rack-style accelerators for AI workloads.
Source: Brother Zhang
Zhang fixes the card by stripping the memory bus
The design in this case appears to match the so-called AFOX RTX 5090 blower card . The problem with these cards is not just the ethical aspect, it is also a post-sales service issue, and apparently there may also be quality concerns. After removing some components, the card’s PCB appeared to be in such poor condition that some traces had to be insulated to avoid short circuits. Honestly, the more the technician showed, the more problems appeared, so it was surprising to see that Zhang still managed to repair the card.
Source: Brother Zhang
The repair itself is unusual because the card had reportedly already been declared dead, with the fault traced to the GPU package and memory lines. The technician found signs of earlier rework, removed the GPU and memory for testing, then restored part of the damaged connections through additional package work. After that, the board was brought back with one memory channel disabled, allowing the card to initialize with 28GB of GDDR7 instead of the full 32GB. And it’s not even 5090 to 5090D, mod; Zhang literally disabled one memory controller.
On one hand, this is an impressive repair, but at the same time, these cards should not exist in the first place. Some will argue that an RTX PRO 6000 costs $10,000 and that this is the only way to get a cheaper card for AI, but it is worth remembering that every one of these cards started out as a gaming model.
Source:
