Vera Rubin Lost 13 Percent of Its Bandwidth, in a Bite

NVIDIA’s Vera Rubin page lists 19.2 TB/s of HBM4 bandwidth per GPU. In September it listed 22.

• Per GPU bandwidth fell 12.7 percent, and NVLink 6 fell from 3.6 to 3 TB/s.
• The 72 GPU rack went from 1,580 to 1,400 TB/s while capacity held at 20.7 TB.
• The eight GPU HGX board lost 26 percent, from 176 to 130 TB/s, and capacity fell from 2.3 TB to 2 TB.
• NVIDIA calls the earlier figures preliminary architectural targets and says the published numbers reflect what the system sustains in volume production.
• The HGX page still prints 22 TB/s per GPU beside a 130 TB/s system total.
• NVIDIA’s own Rubin architecture write-up still claims 2.8 times Blackwell bandwidth, because it uses the old 22 TB/s figure.
• Decode is memory bound, so the cut lands on tokens per second before it lands on training.

AMD’s Instinct MI455X lists 432 GB at 23.3 TB/s per GPU, ahead on both published figures. Read the full analysis.