China is turning modern GeForce RTX 4090 video cards into AI accelerators. What's happening?

At the end of summer 2023 we

wrote about that

that China is exploiting a loophole with cut-down versions of AI accelerators. The fact is that China did not have the opportunity to buy (at least officially) A100 and H100 accelerators. But Nvidia previously released artificially cut-down versions of these systems, the A800 and H800. This was done to circumvent export restrictions.

But a little later, the United States banned the import of any AI chips produced by Nvidia, AMD and Intel into the country. China had to look for new ways to get powerful graphics chips. As far as one can understand, the Celestial Empire has found and is currently actively using this method. Details are under the cut.

Sanctions? What sanctions?

The companies listed above cannot import graphics chips intended for the AI industry into China. The fact is that they use American technology. Accordingly, are required to comply with export regulations established by US regulators. In general, companies do this, but there are also small tricks that allow them to continue supplying modern equipment to buyers from China.

For example, Nvidia created a stripped-down version of the A100 accelerator called the A800. Its difference from the original was that the bidirectional transfer rate (BTR) was reduced by about a third, which made it possible to bypass the new restrictions.

Well, a little later, a stripped-down version of the H100 appeared, which Nvidia is modifying for Chinese consumers. A customized version of the system is sold under a different name – not H100, but H800.

As far as we know, the version of the system that is supplied to China is artificially “slowed down.” Those. The accelerator has reduced throughput characteristics. So, if the H100 has 300 Gbit/s, then the Chinese version has only 150 Gbit/s.

Not only Nvidia, but also Intel also does not want to lose buyers from the Middle Kingdom. Earlier it became known that Intel Corporation began selling Habana Gaudi 2 accelerators to China. As in the case of systems from Nvidia, they are designed to work with deep learning and inference tasks (ensuring the operation of a pre-trained neural network on the end device). At the same time, the accelerator itself is a system that cannot be supplied to China in its current form, due to restrictions imposed by the United States on this country.

But now all these possibilities have sunk into oblivion, as the United States has tightened sanctions, as a result of which neither full nor reduced AI chips can be supplied to the PRC.

What did the Chinese do?

They started buying a huge number of modern gaming video cards Nvidia GeForce RTX 4090 produced by various companies. But not to create gaming PCs, but to turn graphics adapters into accelerators for artificial intelligence.

The 4090 card was chosen because it is the most advanced graphics adapter in the world. Soon after its release, it became in short supply, and not only because gamers began to take it apart. Rather, because China began to purchase these adapters in almost tons, despite the fact that the cost of one device is approximately $2000.

By the way, now 4090 is also banned in terms of supplies to China. But even before the introduction of this ban, Chinese companies managed to purchase a huge number of video cards. A scheme for customizing such modules was previously developed so that they could be turned into AI accelerators. The Chinese have developed a new scheme for them, so that after modification, 4090 cards no longer occupy 3-4 slots in the block, but only 2. This means that they can be installed in servers.

The work is quite painstaking, since most operations have to be performed manually. The Chinese are disassembling the cards, eliminating the cooling system, and then the main components. To create an AI accelerator, a special board has been developed to which these components are transferred. The final product works perfectly in the servers, doing the work the PRC needs.

The whole process is quite complex, almost all stages are carried out by people, so you have to carefully check the functionality of the adapters. The Chinese do this very carefully. In addition to specialized software like Furmark, cards are also tested in artificial intelligence applications. If everything is fine with the cards, they are sent to Chinese companies that develop AI products.

The hybrid graphics adapter is purchased by Chinese data center operators and companies that produce solutions for the AI industry. The country’s domestic market is indeed very large, so it really makes sense for companies that remanufacture new graphics adapters to do such work.

Well, the Chinese also sell the base, the board without the 4090 chip and a couple of other components, only for spare parts. Service centers willingly purchase this kind of thing, because if the video card board fails (physical impact, serious burnout of power connectors, etc.), it can be restored using a donor – the same “bare” board from 4090.

Okay, but what about the cards themselves?

As mentioned above, they quickly became scarce. But now the United States has banned companies from supplying them to China, so market players are hoping for a quick restoration of the supply/demand balance. During the relatively long period of time, the Chinese created such a rush of demand that the already not at all low price of 4090 cards rose very high. Well, the cards themselves have become scarce.

Experts hope that after the ban on the supply of adapters from the Chinese comes into effect, demand will quickly return to normal and prices will fall.