NVIDIA has officially reduced key specifications for its upcoming 'Vera Rubin' AI platform. The company lowered HBM4 memory speed from 11.7 Gbps to 9.6 Gbps per pin, cutting total memory bandwidth in the NVL72 rack to 1,400 TB/s. Additionally, NVLink bandwidth per GPU was reduced from 1.8 TB/s to 1.5 TB/s, affecting bidirectional throughput.
These changes stem from signal integrity challenges, prompting NVIDIA to lower the NVLink Nyquist frequency from 56 GHz to 47 GHz to improve manufacturing yields and system stability. The revised specs place Vera Rubin behind AMD’s Helios system in memory capacity and bandwidth. At the rack level, Helios offers 31 TB of HBM4 and 1.67 PB/s bandwidth, compared to Rubin’s 20.7 TB and 1.4 PB/s.
AMD’s UALoE interconnect also outperforms NVIDIA’s NVLink in aggregate bandwidth.
Source: TechPowerUp News · Summarized by HeadlinesBriefing