Data Localization and Its Importance for Next-Gen Chips Like Rubin

x.com2 min read

NVIDIA's Rubin and Blackwell chips use multi-die packaging, which introduces Non-Uniform Memory Access effects and bandwidth bottlenecks when data crosses the on-chip interconnect known as the stitch. The author explains that data localization—keeping data physically close to the compute units in local High Bandwidth Memory and L2 cache—improves performance and energy efficiency for both memory-bound and compute-bound kernels. CUDA 13.5 addresses this by exposing each die as a locality domain for explicit buffer allocation. Open article →