China’s metamorphosis of gaming into AI: factories dismantle Nvidia graphics cards and turn them into their own AI solutions
If navigating in the conditions of sanctions should be described with a more vivid vocabulary in the technological light and the current time, the right words for it could be precisely the Chinese metamorphosis of gaming in AI. An insider from China’s Baidu forum has shared images that show how China is bending over backwards to circumvent US technology trade sanctions by virtually turning Nvidia and other gaming graphics cards into the products it needs to work with artificial intelligence (AI).
Countries that are under the sanctions of others in economic, technological or any other sense often have to “manage”, that is, to bend in various ways in order to somehow maintain their business or progress. Nowadays, China is one of the examples of this, as we can see from the pictures and information that have been leaked on the Internet, describing this phenomenon.
According to them, specialized factories are being opened throughout China that receive deliveries of Nvidia GeForce RTX 4090 graphics cards, which arrived in this country before the US placed them under the export ban for China. These products already support the AI software ecosystem, and inherently require little or no modification to support the latest major language models.
Given that China’s domestic AI segment is longing for products that would help them develop artificial intelligence, factories are currently modifying the aforementioned graphics cards into their own solutions for large AI models. One of the images shows hundreds of Palit RTX 4090 graphics cards shipped to China, and insiders say factories are expecting new shipments. However, the purpose of these graphics cards is not intended for gaming, but for the domestic development of the AI industry segment.
Before the new US export rules went into effect, Nvidia reportedly sent a large number of AD102 GPUs as well as GeForce RTX 4090 graphics to its partners in China as a priority. It is precisely this priority shipping compared to clients from other markets that is potentially the reason that there is currently a current shortage of these products in the rest of the world and that their price rises to over 2000 US dollars.
However, although the software part is ready for repurposing, the practicality of using the gaming design in a server environment remains. This is exactly what these specialized Chinese factories are for.
Each graphics card has a three- or four-slot design, which means it takes up a lot of space and as such is not suitable for a server AI environment. Workers in specialized factories disassemble each one they receive to make it suitable for a server AI environment.
First, the workers disassemble the large coolers, then remove everything, including the GPU and GDDR6X memory from the PCB part, that is, from the main printed circuit board. Insiders, as Wccftech writes, claim that workers have designed specialized “reference” PCBs that give these AD102 GPUs and GDDR6X memory new life.
After that, each graphics card, or rather the rest of it, is equipped with a cooler with two slots that uses a fan cooling design. Although there are different versions of fan coolers, they all have similar specifications. The Fan GPU Cooler is tuned specifically for large server environments where several GPUs need to work as a team and efficiently dissipate heat from there.
After assembling a practically new device, the newly created Nvidia GeForce RTX 4090 AI solutions undergo rigorous testing in laboratories where we can see Furmark, 3DMark and various AI applications that are launched based on their work. This part of the process is essential for workers to ensure that customer requirements are met in the AI field. When everything is checked, the workers pack the modified devices and send them on to Chinese AI companies.
In addition to the totally unconventional way of using graphics cards, this process also caused the Chinese reseller market to be flooded with RTX 4090 coolers and “bare” printed circuit boards. Resellers continue to sell them at very low prices, often for less than 50 US dollars, since the most valuable parts, the AD102 GPU and GDDR6X memory, have already been disassembled.
They can definitely become a nice collection, but the question is where this engineering expertise leads. However, “bare” PCBs may come in handy in the future, especially when it comes to repairing Nvidia RTX 4090 GPUs, especially since this product still occasionally faces issues such as a faulty 12VHPWR connector.