NVIDIA’s flagship GeForce RTX 4090 is starting to rise in price in the US, while in China, the US sanctions have the GPU banned in the country… but what about the RTX 4090s already there? Well, they’re being turned into AI accelerators and sold at a profit, of course.
VIEW GALLERY – 10 IMAGES
We have news of some Chinese companies buying GeForce RTX series graphics cards and turning them into AI accelerators, given the huge limitations that the Biden administration and the US government have put on AI hardware being exported from the United States and into China. AMD, Intel, and NVIDIA are blocked from selling specific AI hardware to China, while the higher-end RTX 4090 on the gaming GPU side of things is also banned because of its high compute power.
An insider has recently posted on the Chinese Baidu forums, saying that specialized factories are being formed across China that were stockpiling the GeForce RTX 4090 shipments before the bans and now turning them into AI accelerators. We’re talking about GeForce RTX 4090 graphics cards from the likes of Palit, ASUS and its ROG Strix cards, and GIGABYTE’s custom GAMING OC variants.
The massive coolers on the GeForce RTX 4090 graphics cards are disassembled, and then virtually everything is removed from the PCB, apart from the AD102 GPU and GDDR6X memory. A specialized “reference” PCB has been designed, which will take the AD102 GPU + GDDR6 memory and then turn them into an AI accelerator for customers who are in need of AI GPUs in China post-ban.
The sometimes 3-slot or even larger 4-slot gaming GPU is then turned into a smaller dual-slot card with a blower-style fan, with the blower-style GPU cooler tuned specifically for huge server environments where multiple GPUs are working together, generating more heat that needs to be removed from the card. This is versus a gaming system, where the RTX 4090 would be cooled fine in most decent cases with good airflow.
Once the new RTX 4090 AI accelerator is born Frankenstein-style, the card is put through its paces on a test bed running Furmark, 3DMark, and a bunch of AI-focused applications to ensure they work 100%… especially with AI customers.
The software ecosystem on the AI side of things is ready to go with the RTX 4090, with not much modification at all needed to support the latest LLMs (Large Language Models) for AI workloads. NVIDIA recently announced bringing TensorRT and TensorRT-LLM to Windows 11 systems, which will make things even easier again.