An interesting spec of graphics card has popped up in China as found by bdsqlsz on X (via Tomshardware). The AI enthusiast discovered a couple of interesting SKUs available to rent at AutoDL, which have cloud computing servers for AI work. These included both an RTX 4090D with 48GB VRAM and an RTX 4080 Super with 32GB. Both of them have doubled the VRAM of their standalone versions, the former of which is already a China-exclusive GPU.
This was a way around the US ban on exports of AI chips to China and a way for Nvidia to still get high-end hardware over there. Now it seems the customers are finding ways to get even more out of these cards and improve their capabilities. Especially when it comes to AI use, the memory capabilities are in greatest need so it’s no surprise they’re putting in new chips onto the board.
Originally the 4090 had 12 2GB GDDR6X memory modules on board, with the 4080 going for 8 of the same modules. This would mean the company would have replaced these with 4GB modules instead, except these don’t exist. Instead, it likely means they are using custom PCBs to have memory on both sides of the board, something Nvidia does for its top-end Titan and professional cards. However this has not yet done so for these models means the Chinese market, which has been circulating these since the end of June for around $685 more than MSRP, is making them themselves.
A cheap price to pay
As bdsqlsz inquired, it’s nothing too expensive to go for. Although the initial cost of the GPUs might be a bit higher, the rent price is actually more reasonable. For a single RTX 4080 Super, it costs $0.03 per hour to rent out. Obviously running any high computing workloads you’re going to want to rent more than one so the costs will likely shoot up, but there is a lot of power behind these if you can afford it.
But if you’re outside of China, you’re unlikely going to be able to get access to the service, as to sign up you need to have a Chinese phone number. So don’t get too excited as you’re going to have to find something closer to home to run, and likely without fancier modded options.
