NVIDIA lowers memory and NVLink targets for Vera Rubin

NVIDIA has reduced memory and interconnect specifications across its Vera Rubin product line. An updated specification sheet lowers HBM4 memory speed from 11.7 Gbps per pin to 9.6 Gbps per pin. Total memory bandwidth for the Vera Rubin NVL72 rack-scale solution now stands at 1,400 TB/s, down from the previously advertised 1,580 TB/s. The memory limitation is attributed to NVIDIA’s implementation of bus width and electrical signaling rather than directly to its suppliers.

NVLink bandwidth per GPU has also fallen from 1.8 TB/s to 1.5 TB/s, reducing bidirectional GPU bandwidth from 3.6 TB/s to 3 TB/s. NVIDIA reportedly lowered the NVLink Nyquist frequency target from 56 GHz to 47 GHz to address signal integrity constraints.

The lower frequency reduces electrical demands on SerDes, cables, and materials, potentially improving manufacturing yields and making the system easier to build reliably. The revised targets prioritize system stability over peak memory and interconnect performance.