NVIDIA trims Rubin memory and NVLink bandwidth in shipping specs

NVIDIA has revised the published specifications of its Rubin AI accelerators as production hardware arrives. Per-GPU memory bandwidth and NVLink scale-up bandwidth are lower than the preliminary figures, while scale-out networking bandwidth is higher. Compute and memory capacity are unchanged. The change was first noticed by outlets comparing NVIDIA's earlier and current product pages, and NVIDIA has not said why, according to ServeTheHome.

What changed

ServeTheHome reports that NVIDIA was initially targeting 22TB/sec of memory bandwidth per Rubin GPU. NVIDIA's current Vera Rubin NVL72 page lists 288 GB of HBM4 and 19.2 TB/s for a single Rubin GPU, 3 TB/s of NVLink bandwidth, and 0.45 TB/s of networking bandwidth. The preliminary values were 3.6TB/sec for NVLink and 0.4TB/sec for networking.

The numbers, and what they measure

The figures apply at several levels, and they are not interchangeable. The NVL72 rack, which holds 72 GPUs, now lists 1,400 TB/s of aggregate memory bandwidth (previously 1580TB/sec), 216 TB/s of NVLink (previously 260TB/sec) and 32.4 TB/s of networking (previously 28.8TB/sec). The eight-GPU HGX Rubin NVL8 is cut further: HBM4 capacity falls from 2.3TB to 2.0TB and bandwidth from 176TB/sec to 130TB/sec, with NVFP4 training dropping from 280 to 270 PFLOPS. Rubin GPU compute figures, such as 50 PFLOPS of NVFP4 inference, did not change.

NVIDIA told the German outlet PC Games Hardware, via Wccftech, that the revised datasheets reflect series production and that the earlier values were development targets. That is NVIDIA's own account and we could not check it independently. The earlier numbers were labelled preliminary, so this is a correction of targets rather than a retraction of a shipped spec.

Why it matters

Memory bandwidth is often the limit for large-model inference, so a lower figure can matter more than unchanged peak FLOPS. NVLink bandwidth governs how well a rack behaves as one large accelerator. The shift toward more scale-out networking leaves the rack-level compute figures intact but changes the balance between in-rack and between-rack traffic. Buyers comparing Rubin with rival rack designs should use the revised numbers.

What to watch

Independent benchmarks of production Rubin systems will show how much the lower bandwidth affects real workloads. Also worth watching are whether NVIDIA explains the change and whether partner systems quote the old or the new figures.