NVIDIA Quietly Lowers Vera Rubin HBM4 and NVLink Bandwidth Specs, While Boosting Networking
NVIDIA appears to have quietly adjusted the published specifications for its next-generation Vera Rubin AI computing platform, reducing several key bandwidth figures while increasing scale-out networking performance. The changes affect the Rubin GPU, Vera Rubin Superchip, Vera Rubin NVL72 rack-scale system, and HGX Rubin NVL8 platform.
The update was not announced publicly. Instead, it was spotted by comparing current NVIDIA specification pages with archived versions from early September 2026. Archived snapshots from September 1 and September 11 show that the revisions were made sometime during that period.
The most notable change is the reduction in HBM4 memory bandwidth for the Rubin GPU. NVIDIA originally listed each Rubin GPU with 288 GB of HBM4 memory and 22 TB/s of bandwidth. The memory capacity remains unchanged, but the bandwidth figure has now been lowered to 19.2 TB/s.
NVLink 6 bandwidth has also been reduced. The Rubin GPU was previously listed with 3.6 TB/s of NVLink bandwidth, but the updated specification now shows 3 TB/s. That represents a drop of around 17%.
These changes carry over into larger Vera Rubin configurations. The Vera Rubin Superchip, which combines multiple components into a higher-performance package, now lists 38.5 TB/s of HBM4 bandwidth instead of 44 TB/s. Its NVLink bandwidth has also been reduced from 7.2 TB/s to 6 TB/s.
For the flagship Vera Rubin NVL72 system, the total HBM4 capacity remains 20.7 TB, but aggregate memory bandwidth has dropped from 1,580 TB/s to 1,400 TB/s. Total NVLink bandwidth has also been lowered from 260 TB/s to 216 TB/s.
The smaller HGX Rubin NVL8 platform sees some of the largest revisions. Its HBM4 capacity has been reduced from 2.3 TB to 2 TB, while total HBM4 bandwidth has fallen from 176 TB/s to 130 TB/s. That is a reduction of about 26%, making it the most significant bandwidth cut among the listed configurations.
Compute performance figures have also shifted slightly for HGX Rubin NVL8. NVFP4 training performance is now listed at 270 PFLOPS instead of 280 PFLOPS, while FP8 and FP6 training performance has moved from 140 PFLOPS to 130 PFLOPS.
However, not every change is a downgrade. NVIDIA has increased networking bandwidth across the Vera Rubin lineup. The Vera Rubin Superchip networking bandwidth has risen from 0.8 TB/s to 0.9 TB/s. The NVL72 system has increased from 28.8 TB/s to 32.4 TB/s, while HGX Rubin NVL8 has gone from 1.6 TB/s to 1.8 TB/s.
This suggests NVIDIA may be placing more emphasis on scale-out performance, where multiple systems communicate across high-speed networks. For massive AI training and inference clusters, faster networking can help improve communication between racks and nodes, especially in workloads that span thousands of GPUs.
Summary of the key revised NVIDIA Vera Rubin specifications:
Rubin GPU:
HBM4 capacity remains 288 GB
HBM4 bandwidth reduced from 22 TB/s to 19.2 TB/s
NVLink 6 bandwidth reduced from 3.6 TB/s to 3 TB/s
Vera Rubin Superchip:
HBM4 bandwidth reduced from 44 TB/s to 38.5 TB/s
NVLink bandwidth reduced from 7.2 TB/s to 6 TB/s
Networking bandwidth increased from 0.8 TB/s to 0.9 TB/s
Vera Rubin NVL72:
HBM4 capacity remains 20.7 TB
HBM4 bandwidth reduced from 1,580 TB/s to 1,400 TB/s
NVLink bandwidth reduced from 260 TB/s to 216 TB/s
Networking bandwidth increased from 28.8 TB/s to 32.4 TB/s
HGX Rubin NVL8:
HBM4 capacity reduced from 2.3 TB to 2 TB
HBM4 bandwidth reduced from 176 TB/s to 130 TB/s
NVLink Switch bandwidth reduced from 28.8 TB/s to 24 TB/s
Networking bandwidth increased from 1.6 TB/s to 1.8 TB/s
NVFP4 training reduced from 280 PFLOPS to 270 PFLOPS
FP8/FP6 training reduced from 140 PFLOPS to 130 PFLOPS
One interesting detail is that NVIDIA’s documentation does not appear to be fully consistent yet. Some product pages now show the revised system-level specifications, while other sections still list the older Rubin GPU figures, including 22 TB/s of HBM4 bandwidth and 3.6 TB/s of NVLink bandwidth. Some regional pages also appear to still display older data.
NVIDIA has not provided an official reason for the specification changes. One possible explanation is that the original HBM4 bandwidth targets may have been difficult for memory suppliers to meet at scale. By lowering the required bandwidth targets, NVIDIA could potentially improve production yields, simplify supply planning, and ensure enough HBM4 availability for large-scale Vera Rubin deployments.
It is also important to note that these are published specification changes, not independent benchmark results. The real-world impact will depend heavily on the type of AI workload being run. Some workloads are highly sensitive to memory bandwidth and interconnect speed, while others may rely more on compute throughput or benefit from improved networking.
Notably, NVIDIA’s headline inference performance for the flagship Rubin GPU appears unchanged. The advertised 50 PFLOPS of sparse NVFP4 inference performance is still listed, suggesting NVIDIA may be confident that the platform can deliver strong AI performance despite the reduced memory and NVLink bandwidth figures.
The Vera Rubin platform remains one of NVIDIA’s most important upcoming AI architectures, designed to succeed the Blackwell generation in high-end data center and AI supercomputing environments. Even with the revised specifications, the platform is expected to offer massive performance, large HBM4 memory capacity, and advanced rack-scale connectivity for next-generation AI training and inference workloads.
For AI data centers, the key question now is how these revised Vera Rubin specifications will affect performance in real deployments. Lower HBM4 and NVLink bandwidth could matter for bandwidth-heavy models, but improved networking may help larger clusters scale more efficiently. Until production systems are tested with real AI workloads, the full impact of NVIDIA’s quiet specification update will remain uncertain.
