NVIDIA scales back memory and NVLink bandwidth for Rubin GPUs

Revised hardware specifications

NVIDIA has officially updated the shipping specifications for its upcoming Rubin GPU architecture. While the company originally promoted more aggressive targets for memory and interconnect speeds, the new figures reflect a slight cooling of those expectations. These changes come as the company prepares for the production ramp-up of the Vera Rubin series in late 2026.

The most notable change involves HBM4 memory bandwidth, the speed at which the GPU can access its high-bandwidth memory. NVIDIA originally targeted 22TB/second of memory bandwidth per Rubin GPU, but the final shipping spec is 19.2TB/second. This represents a roughly 11% decrease in the data rate per pin compared to the preliminary goals.

The company did not provide a specific reason for these revisions. However, industry analysts suggest that the adjustments may stem from difficulties in hitting high clock speeds or achieving sufficient yields for the new HBM4 memory stacks. Despite the reduction, the Rubin GPU still targets a data rate significantly higher than some of its competitors.

Internal vs. external networking

The updates also impact NVLink, which is the technology NVIDIA uses to connect multiple GPUs so they can act as a single unit. This is often called scale-up networking. The bandwidth for this connection on a single Rubin GPU has been lowered from 3.6TB/second to 3.0TB/second, a 17% reduction from the initial plan.

In contrast, NVIDIA is overperforming on scale-out networking, which is the system used to connect different racks or clusters of GPUs. The networking bandwidth for a single Rubin GPU has been increased from 400GB/second to 450GB/second. This suggests a strategic shift toward supporting larger, distributed AI models that require high-speed communication across many nodes.

These adjustments mean that while the internal communication between chips in a single box is slower than hoped, the ability to link those boxes together is stronger. This balance may be intended to optimize for the way hyperscalers deploy massive AI clusters. The theoretical compute throughput for inference and training remains unchanged at 50 PFLOPS and 35 PFLOPS respectively.

Impact on the NVL72 rack

The Vera Rubin NVL72 is NVIDIA's flagship rack-scale system that integrates CPUs, DPUs, and Rubin GPUs. Because it relies on the base Rubin hardware, its specifications have been revised in lockstep with the individual chip changes. The cumulative GPU memory bandwidth for the rack has dropped from 1,580TB/second to 1,400TB/second.

Similarly, the NVLink scale-up bandwidth for the entire rack has been adjusted downward from 260TB/second to 216TB/second. This matches the 17% reduction seen in the individual GPU units. However, the rack still benefits from the improved scale-out networking, which was boosted from 28.8TB/second to 32.4TB/second.

While the total memory capacity for the NVL72 remains at 20.7TB of HBM4, the lower bandwidth may affect certain high-intensity workloads. It remains unclear if the system can hit the exact same performance levels as originally projected. Nevertheless, the GPU compute hardware and the Vera CPUs are still shipping according to their original performance targets.

The HGX NVL8 memory cut

The HGX NVL8 systems are seeing even more significant changes than the rest of the Rubin family. In addition to the standard bandwidth reductions, NVIDIA is cutting the total amount of HBM4 memory available in these systems. The memory capacity is dropping from a planned 2.3TB down to a flat 2.0TB.

This reduction means that each of the eight GPUs in an NVL8 system will only use seven of its eight memory stacks. Because of this, the NVL8 is taking a much larger hit to its memory bandwidth than other products. It is losing about 26% of its memory bandwidth, falling to 16.8TB/second per GPU.

These changes result in a slight decrease in training performance for the NVL8 platform. FP4 training performance has been dialed down from 280 PFLOPS to 270, and FP8 training has dropped from 140 PFLOPS to 130. Despite these cuts, the inference performance for the system remains steady at 400 PFLOPS for FP4.

What this means for the market

NVIDIA's decision to revise these specs suggests a pragmatic approach to hardware manufacturing. By adjusting the targets now, the company ensures that the hardware it ships to customers will be reliable and consistent. It is common for server hardware specifications to shift during the long lead times between announcement and production.

For customers, the impact depends heavily on the specific AI workload being performed. Some models may be bottlenecked by memory speed, while others may benefit more from the increased scale-out networking. Since these changes apply to all shipping Rubin hardware from day one, the industry will have to adapt to these new performance baselines.

As the Vera Rubin ramp-up continues into the fourth quarter of 2026, these specs will define the next generation of AI infrastructure. While the hardware is slightly less powerful than the initial marketing suggested, it represents the finalized production reality for NVIDIA's next major leap in GPU technology.