The NVIDIA RTX 4090D graphics card, aimed at the Chinese market and based on the AD102 processor, has been developed to circumvent U.S. export restrictions, and is now available with double the memory capacity — 48 GB.
The information about these new graphics cards comes from insider @bdsqlsz, who gained access to a cloud computing platform where these devices are already deployed. The primary use for these graphics cards is training large language and other generative AI models that not only require significant computational power but also a substantial amount of memory.
While the original tweet does not include photos or physical evidence, it confirms the memory bandwidth of the RTX 4090D at 937 GB/s, matching GDDR6X memory speeds. The 48 GB capacity is paired with the AD102 GPU, typically found in RTX ADA workstations, but utilizes the GDDR6 standard. This indicates that the 48 GB RTX 4090D features faster memory. Additionally, the software references the GeForce RTX 4080 SUPER with 32 GB, which usually comes with 16 GB of GDDR6X memory.
Such memory modifications are not uncommon, especially in the modding community and GPU service centers, where memory replacement is a routine task. However, the fact that the cloud computing platform offers such modified graphics card versions speaks to their more systematic use.
The RTX 4090D typically supports only 12 memory modules, but it can utilize RTX 3090 Ti boards, which support 24 modules, alongside the AD102 GPU from the RTX 4090D. This solution effectively doubles the memory capacity. It is likely that Chinese servers implement such modifications.
Sources: Tom's Hardware, VideoCardz
Comments (0)
There are no comments for now