The enormous demand for AI is forcing fundamental changes across every layer of the technology stack. When transmitting data between racks, computer manufacturers are increasingly turning to silicon photonics technology to overcome the limitations of copper. This includes NVIDIA, which has made significant investments in this field. However, many of these decisions are made out of necessity.
At the recent GTC developer conference, NVIDIA CEO Jensen Huang briefly outlined the company’s plans to build the NVL1152 “Kyber” supercomputer. This is an entirely new, massive-scale-up supercomputer that will feature 1,152 GPUs. The NVL1152 is expected to use the “Feynman” generation GPU, currently planned for shipment in 2028. The number of processors will be 16 times that of the company’s current largest scale-up supercomputer, the NVL72.