NVIDIA Cuts Vera Rubin HBM4 and NVLink Bandwidth Targets

By Amara Osei

NVIDIA has cut the HBM4 speed target for its Vera Rubin platform from 11.7 Gbps per pin to 9.6 Gbps, reducing listed memory bandwidth for the 72-GPU NVL72 rack to 1,400 TB/s from 1,580 TB/s.

What changed in the revised specification?

The changes cover NVIDIA's Vera Rubin product line and affect both memory and the NVLink interconnect fabric. Its updated product specification sheet lowers per-GPU NVLink bandwidth from 1.8 TB/s to 1.5 TB/s. Measured bidirectionally, the target falls from 3.6 TB/s to 3 TB/s for each GPU.

The report attributes the memory reduction not to NVIDIA's suppliers directly, but to the company's implementation of memory bus width and electrical signaling. It says those elements cannot support the higher speed previously advertised.

NVIDIA also reportedly lowered the NVLink Nyquist-frequency target from 56 GHz to 47 GHz. A lower frequency reduces electrical demands on the SerDes, cables and materials, which can improve manufacturing yields and overall feasibility.

The trade-off is explicit: NVIDIA gives up some peak bandwidth to improve system stability and manufacturability. The comparison graphic shows the revised specification in light mode and the previous specification in dark mode.

The revised specification lowers both Rubin's memory and interconnect peaks.

How does Rubin compare with AMD Helios?

On the figures cited, AMD's rack-scale Helios system exceeds the revised Vera Rubin in both HBM4 capacity and memory bandwidth. The Instinct MI455X carries 432 GB of HBM4 per GPU at 23.3 TB/s, compared with 288 GB at 19.2 TB/s for a Rubin GPU.

At the 72-GPU rack level, Helios provides roughly 31 TB of HBM4 and about 1.67 PB/s of aggregate memory bandwidth. Vera Rubin NVL72 provides approximately 20.7 TB and 1.4 PB/s.

The comparison uses the revised Rubin numbers, including the lower 9.6 Gbps-per-pin HBM4 target. Its 1.4 PB/s rack bandwidth matches the new 1,400 TB/s figure listed for NVL72.

Helios leads Rubin on the cited on-package capacity, rack memory bandwidth and per-GPU bandwidth.

What changes in the scale-up fabric?

AMD's UALoE scale-up interconnect also posts higher figures. It delivers 3.6 TB/s of bidirectional bandwidth per GPU and 260 TB/s in aggregate. NVIDIA's revised NVLink target is 3 TB/s per GPU, or 216 TB/s across the 72-GPU rack.

For NVIDIA, the reduction starts with a 1.8 TB/s per-GPU NVLink figure and a 3.6 TB/s bidirectional figure. The new sheet replaces those values with 1.5 TB/s and 3 TB/s, respectively. AMD's 3.6 TB/s bidirectional per-GPU rate therefore matches Rubin's former target rather than its revised one.

AMD also leads on the cited scale-up bandwidth at both GPU and rack levels.

What does the specification change mean?

Rubin's revised profile now combines three lower figures: 9.6 Gbps per HBM4 pin, 1.4 PB/s of memory bandwidth across NVL72 and 3 TB/s of bidirectional NVLink bandwidth per GPU. Those replace 11.7 Gbps, 1,580 TB/s and 3.6 TB/s.

The report frames the change as an engineering compromise rather than a direct supplier problem. Reducing the Nyquist target eases signal-integrity and manufacturing demands across SerDes, cables and materials, while the lower HBM4 target reflects bus-width and signaling constraints.

AMD, by comparison, lists 31 TB of HBM4 and 1.67 PB/s of memory bandwidth for a 72-GPU Helios rack, plus 260 TB/s of aggregate UALoE bandwidth. Rubin's corresponding totals are about 20.7 TB, 1.4 PB/s and 216 TB/s.

NVIDIA has traded peak throughput for manufacturability, while Helios retains the higher cited memory and link figures.