Login to Continue Learning
Factories in China have been producing custom AI GPUs based on the NVIDIA GeForce RTX lineup since the rise of artificial intelligence a few years ago. These factories create cards like the RTX 4090 with 48 GB and even up to 96 GB VRAM (though this was one-off). The latest model to get the custom memory treatment is the RTX 5090, which originally comes with 32 GB of GDDR7 memory.
These factories achieve such modifications by salvaging GPUs and memory from gaming graphics cards like the GeForce RTX 5090. They then equip these components on custom PCBs that often allow for dual-side memory configurations. This allows them to create custom graphics cards with much larger capacities, not just through simple component swaps but also using proper firmware and BIOS.
In addition to memory modding, gaming GPUs are converted into two-slot blower-style designs for easier use in server and workstation setups.
According to I_Leak_VN, the latest model to receive this treatment is the RTX 5090. While the standard card features 32 GB of GDDR7 memory, the custom variant has a massive 128 GB frame buffer, quadrupling its memory capacity. It appears that the card uses higher-capacity GDDR7 memory dies, likely resulting in 64 GB on one side and another 64 GB on the other side.
The RTX 5090 with 128 GB is reportedly priced at around $13,200 US, which is a 6.6x higher price for 4x the memory capacity compared to the standard card’s retail price of around $2500-$3000.
While the price is steep, it aligns with other high-capacity professional GPUs like the RTX PRO 6000 Blackwell, which costs around $10,000 US. The 128 GB GeForce RTX 5090 offers 33% more memory and 33% higher price than the RTX PRO 6000.
The custom mod does not have any approval from NVIDIA and is sold by these factories to AI customers since such a large amount of VRAM has no practical use in gaming.
