GeForce RTX 2080 Ti unexpectedly become a hit for local AI - cards with doubled memory sold for $500

The flagship GeForce RTX 2080 Ti from the Turing generation, released nearly eight years ago, originally came with 11GB of GDDR6 memory. Now, unofficially modified versions with doubled VRAM are actively sold on secondary markets as ready-made solutions for local AI and compute tasks. Sellers on eBay offer these modified cards with 22GB of GDDR6 for prices starting at $499, according to VideoCardz. Interestingly, sellers do not guarantee a specific brand; possible brands include Gigabyte, MSI, Asus, and Leadtek, depending on stock availability. The modification involves replacing 1GB GDDR6 memory chips with 2GB modules, and since the RTX 2080 Ti supports 11 active memory channels over a 352-bit bus, the total VRAM doubles to 22GB. This also requires adjusting the card's memory configuration so the GPU recognizes the larger modules. While the extra memory does not boost the raw performance of the TU102 GPU with 4352 CUDA cores, it allows larger AI models, datasets, and image generation workloads to fit entirely in local memory, making these old Turing cards an unusual alternative for users prioritizing memory capacity over current-generation performance. The RTX 2080 Ti also retains CUDA and Tensor Core support, although its tensor cores do not support some lower-precision formats and other features introduced in newer Nvidia architectures. One seller claims these modified cards work with standard Nvidia drivers without custom software.