The rapid increase in demand for computing power in the field of artificial intelligence (AI) has led to innovative solutions in China. To meet the growing demands, Chinese companies are modifying NVIDIA’s GeForce RTX 40 series graphics processing units (GPUs) and equipping them with expanded video memory (VRAM). These modifications are particularly popular with cloud computing providers who rent high-performance hardware for AI applications to their customers.
A striking example of these modifications is the NVIDIA GeForce RTX 4090D, whose VRAM has been expanded to 48 GB GDDR6X memory. These GPUs are rented out by Chinese cloud service providers for just USD 0.03 per hour. Another modified GPU, the GeForce RTX 4080 SUPER with 32 GB of memory, is also available for both rental and purchase. Modifying GPU VRAM is not a new practice. Such modifications have been made in the past to increase performance – often at comparatively low cost. One example of this is the NVIDIA GeForce RTX 2080 Ti, whose VRAM was doubled to 22 GB by modders to meet the requirements of the AI market. Now this modding practice also seems to have reached the latest GeForce RTX 40 “Ada Lovelace” GPUs.
According to a Chinese AI expert, modified GeForce RTX 4090D and RTX 4080 SUPER GPUs are circulating in the region. These GPUs are particularly in demand from cloud computing providers who rent out their computing power to meet the increasing demand in the market. The modified GeForce RTX 4080 SUPER, which can be rented for just USD 0.03 per hour, offers remarkable value for money. The modified GeForce RTX 4090D 48 GB is also an interesting option for purchase: it is reportedly available on the domestic market for around 2,500 US dollars. In comparison, the RTX 6000 “Ada”, which also offers 48 GB of memory, costs around 7,000 US dollars, making the modified versions attractively priced.
Exactly how these modified GPUs are created remains unclear, as such operations are carried out in secret. So far, no images of the modified models have been published. However, it is believed that the modders were able to design a PCB that can accommodate 24 GDDR6X modules on the GeForce RTX 4090D. In the past, Chinese factories have been observed disassembling NVIDIA gaming GPUs to fit them with new PCBs and cooling systems better tailored to large-scale AI operations. These measures were taken after the US government announced it would restrict the export of AI accelerators from NVIDIA to China. It would be interesting to see what the modified GPUs look like in detail, especially the solutions to the challenges that come with such a large VRAM on a GPU. The cooling system that has to support these powerful components should also be of particular interest.
The fact that these modified GPUs are not available outside of China, as access to the applications requires a local number and identification, underlines the exclusivity of these offerings. Nevertheless, China is once again demonstrating its innovative strength in the technological field, particularly by adapting and modifying existing hardware for specialized applications such as AI development. China thus remains a pioneer when it comes to finding new ways to meet growing technological demands, be it through modding capabilities or other technological developments.
Source: 青龍聖者 via X
58 Antworten
Kommentar
Lade neue Kommentare
Veteran
Mitglied
Veteran
Mitglied
Veteran
Mitglied
Urgestein
Urgestein
Mitglied
Mitglied
Veteran
Veteran
Urgestein
Urgestein
Urgestein
Urgestein
Mitglied
Veteran
Veteran
Alle Kommentare lesen unter igor´sLAB Community →