NVIDIA GeForce RTX 4090D with 48GB and RTX 4080 SUPER 32GB now offered in China for cloud computing

RTX 4090D and RTX 4080 SUPER get doubled memory

Where there is demand, there are custom solutions available.

The China-centric RTX 4090D graphics card, designed to circumvent U.S. export restrictions on the flagship AD102 gaming GPU, is now available with doubled memory size of 48GB. This information was first reported by @bdsqlsz, who accessed a cloud computing platform where these instances were already installed.

The primary use of these custom graphics cards is not for cryptomining but for training large language models and other generative AI models, which require significant computational power and memory capacity.

RTX4090D/4080S mods, Source: @bdsqlsz

Although the original tweet does not include a photograph or physical evidence, it confirms the RTX 4090D’s memory bandwidth at 937 GB/s, consistent with GDDR6X memory speeds. The 48GB memory capacity is available with the AD102 GPU, typically seen in the RTX ADA workstation series, but uses the GDDR6 standard. This suggests that the 48GB RTX 4090D has faster memory.

Additionally, the software output lists a GeForce RTX 4080 SUPER , which typically features 16GB of GDDR6X memory. However, it seems the cloud computing service offers custom 4080 cards with 32GB of memory installed.

Memory modifications like these are not uncommon, especially within the modding community and GPU servicing centers, where memory replacements are a daily task. However, the fact that a cloud computing platform offers instances of such modified graphics cards suggests a larger-scale supply and the use of custom board designs.

The RTX 4090D typically supports only 12 memory modules, but it is possible to use the PCB from an RTX 3090 Ti, which supports 24 modules, along with the AD102 GPU from the RTX 4090D. This solution allows for doubling the memory capacity. In other words, it appears that a company is using a custom RTX 3090 Ti PCB with an AD102 GPU to achieve this upgrade.

Source: @bdsqlsz via Tom’s Hardware