Nvidia's shipping Rubin specs cut HBM4 bandwidth to 19.2TB/s and NVLink to 3.0TB/s per GPU from earlier targets
ServeTheHome details the revised datasheet as first production hardware arrives: memory bandwidth falls from 22TB/s (about 11%), NVLink from 3.6TB/s, scale-out networking rises to 450GB/s, and HGX Rubin NVL8 drops from 2.3TB to 2.0TB of HBM4; compute figures are unchanged, Nvidia told Wccftech the new numbers reflect production systems, and it has not said why bandwidth was lowered.