According to reports, NVIDIA originally planned to use four GPU compute chiplets in the Rubin Ultra AI accelerator, set to be launched in 2027, to enhance performance. However, due to multiple challenges in the manufacturing process, including packaging yield, thermal management feasibility, and overall manufacturing costs, NVIDIA has canceled the four-compute-chiplet version and instead adopted a more mass-production-friendly dual-GPU design. Despite halving the number of chips, NVIDIA will strive to maintain overall performance by upgrading HBM memory specifications and making architectural tweaks.
