By Karsten Rabeneck-Ketme
Publication Date: 2026-11-02 05:00:00
When introducing new AI accelerators, manufacturers regularly outdo one another with impressive performance claims. However, things only get interesting when the hardware presented actually becomes available for purchase. NVIDIA is now at precisely this point with its upcoming Vera Rubin platform. Updated technical documents surfaced in late October 2026, with some details differing from figures published earlier. Memory bandwidth and certain interconnect specifications are particularly affected. For operators of large AI data centers, such differences are by no means trivial, since memory accesses and communication between GPUs often determine how efficiently large language models can be trained and run.
With Vera Rubin, NVIDIA is preparing the next generation of its AI data center platform. At its heart is the NVL72 system, which combines a total of 72 Rubin GPUs and 36 Vera CPUs within a shared rack architecture. The accelerators are connected via NVLink, allowing large volumes of data to be exchanged between individual processing units. ConnectX-9 network adapters and BlueField-4 DPUs are also included for communication with other systems and to handle infrastructure tasks.
According to a ServeTheHome report dated October 31, 2026, the latest product documents show changes from specifications communicated earlier. While NVIDIA initially cited HBM4 memory bandwidth of up to 22 TB/s per GPU, updated figures point to…
