Chinese Companies Double NVIDIA RTX 40 Series VRAM for AI Workloads

NVIDIA‘s popular GeForce RTX 40 series graphics cards are getting a significant upgrade in China, where companies are doubling their VRAM to enhance performance for AI workloads. These modified GPUs, including the GeForce RTX 4090D with 48GB of memory and the GeForce RTX 4080 Super with 32GB, are being offered by Chinese cloud computing providers like AutoDL. These enhanced GPUs are available for rent at incredibly low prices, such as $0.03 per hour for the RTX 4080 Super.

Increased Memory for AI Tasks

The modification process involves replacing the existing GDDR6X memory modules with higher-capacity ones or using a custom PCB to fit more memory modules. This approach mirrors techniques used in NVIDIA’s more premium models, like the Titan and RTX professional cards. The additional memory is particularly useful for AI tasks, where more VRAM can significantly improve performance.

While these modified GPUs are mainly used within China, the global tech community is taking note. These GPUs offer a cost-effective alternative to more expensive AI-focused hardware, like the RTX 6000 Ada, which also features 48GB of memory but costs around $7,000.

The modifications highlight China’s innovative approach to circumventing restrictions on high-performance AI hardware exports.