NVIDIA has cut several listed bandwidth figures for its Vera Rubin AI systems, while some network speeds have gone up. A check of older and newer product documents shows the changes across several hardware setups. But these are spec sheet revisions, not test results from machines in use.
For the Vera Rubin Superchip, HBM4 memory bandwidth went from 44 TB/s to 38.5 TB/s. NVLink also dropped from 7.2 TB/s to 6 TB/s. Meanwhile, the scale-out network figure rose from 0.8 TB/s to 0.9 TB/s.
And this is not just one chip. The Vera Rubin NVL4 and NVL72 setups show a similar trend in the reported files. NVL4 has one key twist, though. Its scale-out network rate jumped from 0.4 TB/s to 1 TB/s, a big gain on paper.
The HGX NVL8 has more cuts in its listed specs. Its HBM memory went from 2.3 TB to 2 TB, while total memory bandwidth fell from 176 TB/s to 130 TB/s. The NVLink Switch rate also went from 28.8 TB/s to 24 TB/s.
Elsewhere in that HGX setup, the listed NVFP4 and FP8/FP6 training figures fell as well. Yet two-way network bandwidth climbed from 1.6 TB/s to 1.8 TB/s. The updates are not all cuts, and each figure covers a different part of the system.
NVIDIA's live specs back up the new Superchip numbers. Its NVL72 page also gives 20.7 TB of HBM4 and a total memory rate of 1,400 TB/s for a full rack. Those rack figures should not be mixed up with the rates for one chip.
There is a small but key change for the Vera CPU as well. Older material gave its LPDDR5X memory as 1.5 TB per CPU. The newer text treats 1.5 TB as the top limit, rather than a fixed amount for each unit.
As for why NVIDIA made these edits, there is no firm answer yet. One theory is that a lower HBM speed target could let the firm use a wider range of memory parts, which may help with supply. Still, NVIDIA has not confirmed that reason.
For buyers, the takeaway is fairly simple. Check the latest specs for the exact model and setup before judging its speed or value. None of these document changes prove a real-world drop in performance, a launch delay, or a broken product pledge.
For the Vera Rubin Superchip, HBM4 memory bandwidth went from 44 TB/s to 38.5 TB/s. NVLink also dropped from 7.2 TB/s to 6 TB/s. Meanwhile, the scale-out network figure rose from 0.8 TB/s to 0.9 TB/s.
And this is not just one chip. The Vera Rubin NVL4 and NVL72 setups show a similar trend in the reported files. NVL4 has one key twist, though. Its scale-out network rate jumped from 0.4 TB/s to 1 TB/s, a big gain on paper.
The HGX NVL8 has more cuts in its listed specs. Its HBM memory went from 2.3 TB to 2 TB, while total memory bandwidth fell from 176 TB/s to 130 TB/s. The NVLink Switch rate also went from 28.8 TB/s to 24 TB/s.
Elsewhere in that HGX setup, the listed NVFP4 and FP8/FP6 training figures fell as well. Yet two-way network bandwidth climbed from 1.6 TB/s to 1.8 TB/s. The updates are not all cuts, and each figure covers a different part of the system.
NVIDIA's live specs back up the new Superchip numbers. Its NVL72 page also gives 20.7 TB of HBM4 and a total memory rate of 1,400 TB/s for a full rack. Those rack figures should not be mixed up with the rates for one chip.
There is a small but key change for the Vera CPU as well. Older material gave its LPDDR5X memory as 1.5 TB per CPU. The newer text treats 1.5 TB as the top limit, rather than a fixed amount for each unit.
As for why NVIDIA made these edits, there is no firm answer yet. One theory is that a lower HBM speed target could let the firm use a wider range of memory parts, which may help with supply. Still, NVIDIA has not confirmed that reason.
For buyers, the takeaway is fairly simple. Check the latest specs for the exact model and setup before judging its speed or value. None of these document changes prove a real-world drop in performance, a launch delay, or a broken product pledge.
