NVIDIA lowers some Rubin GPU bandwidth specs, boosts network speed

A comparison of NVIDIA's latest documents with earlier versions shows some spec changes for the Rubin GPU system. The Vera Rubin superchip's HBM4 memory bandwidth drops from 44TB/s to 38.5TB/s, and NVLink bandwidth falls from 7.2TB/s to 6TB/s. But scale-out network bandwidth rises from 0.8TB/s to 0.9TB/s, and Vera Rubin NVL4 gets a boost to 1TB/s. For the Rubin GPU SMX module, two HGX NVL8 systems see HBM memory cut from 2.3TB to 2TB, with bandwidth down from 176TB/s to 130TB/s. Training compute for NVFP4 and FP8/FP6 is lower too, and NVLink Switch bandwidth drops to 24TB/s, while bidirectional network bandwidth goes up to 1.8TB/s. The lower bandwidth bar may let more HBM4 memory qualify for Rubin GPUs, which could improve production yields. Vera CPU memory is now listed as up to 1.5TB of LPDDR5X.