In brief
- The Rubin GPU’s HBM4 capacity remains 288 GB, while memory bandwidth has been revised down to 19.2 TB/s.
- NVLink bandwidth per GPU has fallen from 3.6 TB/s to 3 TB/s; network bandwidth has increased from 400 GB/s to 450 GB/s.
- On Vera Rubin NVL72, memory bandwidth has dropped from 1,580 TB/s to 1,400 TB/s, and NVLink bandwidth from 260 TB/s to 216 TB/s.
- HGX Rubin NVL8 is now listed with 2 TB of HBM4 instead of 2.3 TB, and 270 PFLOPS of FP4 training performance instead of 280 PFLOPS.
NVIDIA has updated the shipping specifications for GPUs and systems in its Vera Rubin family. The changes were published on the company’s product pages and architecture datasheet in September. According to an October 31 report by ServeTheHome, Rubin GPU memory and NVLink bandwidth have been reduced from their initially announced figures, while network bandwidth has increased. The updates are also reflected in systems such as Vera Rubin NVL72 and HGX Rubin NVL8.
What changed in the Rubin GPU specs?
The HBM4 capacity of a single Rubin GPU remains 288 GB. However, memory bandwidth has been reduced from the initially stated 22 TB/s to 19.2 TB/s. NVLink scaling bandwidth is now listed at 3 TB/s instead of 3.6 TB/s. In contrast, scale-out network bandwidth has increased from 400 GB/s to 450 GB/s.
Some key compute performance figures remain unchanged: NVIDIA’s stated NVFP4 inference performance is still 50 PFLOPS, and training performance remains 35 PFLOPS. The company has not explained the reason for the spec changes. ServeTheHome points to possibilities such as HBM4 memory speeds or manufacturing yields, but NVIDIA has not confirmed these as the reasons.
How are the NVL72 and NVL8 systems affected?
On Vera Rubin NVL72, total HBM4 capacity remains 20.7 TB, while memory bandwidth has fallen from 1,580 TB/s to 1,400 TB/s. Total system NVLink bandwidth has also dropped from 260 TB/s to 216 TB/s. Network bandwidth, meanwhile, has risen from 28.8 TB/s to 32.4 TB/s.
The changes are more extensive for HGX Rubin NVL8. HBM4 capacity has been reduced from 2.3 TB to 2 TB, and memory bandwidth from 176 TB/s to 130 TB/s. NVLink is now listed at 24 TB/s instead of 28.8 TB/s, while network bandwidth has increased from 1.6 TB/s to 1.8 TB/s. NVFP4 inference performance remains at 400 PFLOPS; training performance has been lowered from 280 PFLOPS to 270 PFLOPS, and FP8 training performance from 140 PFLOPS to 130 PFLOPS. According to the report, these changes also apply to NVL8-based HGX and DGX systems.
What this means for architecture and visualization workflows
These products are systems designed for AI and data center infrastructure, not direct updates to architectural modeling or visualization software. As a result, the revised figures do not, by themselves, mean a direct performance change in an architecture firm’s day-to-day 3D workflow. However, for teams using AI services running on large GPU clusters, memory bandwidth, inter-GPU communication, and network connectivity are metrics that need to be considered separately.
How a workload is affected by these figures depends on how the system is used. The source provides no information about product pricing, licensing terms, availability in Turkey, or compatibility with specific architecture software. Any decision therefore requires verified information not only about theoretical compute figures, but also about the memory and connectivity requirements of the intended workload.
Sources
1 sourceSource texts are not republished; short quotes are marked, everything else is our own summary and commentary.
The Rubin updates show that high compute figures alone do not tell the whole story about system performance. Memory and inter-GPU connectivity should also factor into purchasing decisions for large-scale AI workloads. The capacity and bandwidth changes on NVL8, in particular, should be assessed together.
For architecture firms and visualization teams in Turkey, however, this news is not in itself a reason to upgrade hardware. The source does not specify the price, licensing, or compatibility of these systems with architecture software. Investment decisions should not be based on the product family name alone before the actual hardware configuration of the service being used and its impact on the workload are clear.
Frequently asked questions
Has the NVIDIA Rubin GPU’s HBM4 capacity changed?
A single Rubin GPU still has 288 GB of HBM4. The revised figure is for memory bandwidth, not capacity: it is now 19.2 TB/s instead of 22 TB/s.
What are the updated bandwidth figures for Vera Rubin NVL72?
NVL72’s memory bandwidth has been updated to 1,400 TB/s and NVLink bandwidth to 216 TB/s. Network bandwidth has increased to 32.4 TB/s.
Which HGX Rubin NVL8 specs have been reduced?
HGX Rubin NVL8’s HBM4 capacity has been reduced to 2 TB, memory bandwidth to 130 TB/s, and NVLink bandwidth to 24 TB/s. Its FP4 training figure has also been updated to 270 PFLOPS.
