NVIDIA Updates Rubin GPU Specs as First Production Hardware Arrives

As we enter the fourth quarter of 2026, NVIDIA and its partners are ramping up production of the Vera Rubin systems, with hyperscalers beginning to make NVL72 systems available to customers. With this production rollout, NVIDIA has finalized hardware specifications for the Rubin GPU, revealing several key changes from their preliminary announcements.

According to updated specifications first spotted by IT Home and posted on NVIDIA's webpages in September, the Rubin GPU now offers less memory and scale-up networking bandwidth than previously promoted, but more scale-out bandwidth. While specification changes before product launches are common, NVIDIA's case involves significant reductions in key performance metrics.

Memory Bandwidth Reduction

The most notable change is in HBM4 memory bandwidth. NVIDIA initially targeted 22TB/second per Rubin GPU, but the final specification provides 19.2TB/second—a reduction of 1.375Gbps per pin or approximately 11%. This change comes despite NVIDIA's aggressive data rate targets, which were significantly higher than the official HBM4 specification and even higher than rival AMD's planned 7.6Gbps/pin for their MI455X.

The reduction suggests NVIDIA and its memory partners may have faced challenges with memory controller speeds, stack speeds, or yields. While NVIDIA hasn't commented on the reasons for this change, the result is less GPU memory bandwidth for customers, potentially affecting performance for certain workloads.

NVLink and Network Bandwidth Adjustments

NVLink interconnect bandwidth has also been reduced from 3.6TB/second to 3.0TB/second per GPU—a 17% decrease. This is particularly significant as NVLink enables multiple GPUs to function as a single unit, making it a critical component for NVIDIA's high-performance computing solutions.

However, scale-out networking bandwidth has increased from 400GB/second to 450GB/second per GPU, providing a performance boost for larger deployments that rely heavily on scale-out networking. The theoretical compute throughput remains unchanged at 50 PFLOPS for FP4 inference and 35 PFLOPS for FP4 training.

Product Line Impacts

These specification changes affect all Rubin-based products. For NVIDIA's flagship Vera Rubin NVL72 rack, cumulative GPU memory bandwidth has decreased from 1,580TB/second to 1,400TB/second (11% reduction), while NVLink bandwidth across the system has dropped from 260TB/second to 216TB/second (17% reduction). Scale-out networking bandwidth for the rack increased from 28.8TB/second to 32.4TB/second.

The HGX Rubin NVL8 has been impacted most severely, with additional reductions in HBM4 memory capacity from 2.3TB to 2.0TB (13% reduction). This results in each GPU using only 7 of 8 memory stacks, causing a 26% reduction in memory bandwidth for these systems. Training performance for FP4 and FP8 has been reduced accordingly, though inference performance remains unchanged at 400 PFLOPS for FP4.

Market Implications

These changes suggest NVIDIA has had to temper its ambitions for the Rubin GPU to bring the product to market at scale. The full performance impact remains unclear, as not all workloads are equally affected by memory or networking bandwidth limitations. Despite these adjustments, NVIDIA is proceeding with production, with significant quantities of Rubin hardware expected to ship over the coming quarters.