NVIDIA sets HBM4 bandwidth at 19.2 terabytes per second per Rubin GPU
Key Facts
NVIDIA has updated the published specifications for its Vera Rubin platform, setting HBM4 memory bandwidth at 19.2 terabytes per second per Rubin GPU, compared with a previous development target of 22 terabytes per second. The company's current page lists memory capacity of 288 gigabytes per unit. A NVIDIA representative said, in a response cited by Wccftech, that specifications published during development were "initial architectural targets." The representative added that the final specifications reflect what the system sustains in qualified large-scale production. This wording places the change in the context of a shift from development targets to production-system specifications.
HBM4 capacity per Rubin GPU remains at 288 gigabytes in the before-and-after comparison table, according to the report citing NVIDIA's specifications. The bandwidth update therefore differs from memory capacity per unit, which the table described as "unchanged." The report also said NVLink Switch bandwidth for the HGX Rubin NVL8 system is now 24 terabytes per second, down from 28.8 terabytes per second. These figures refer to the NVLink Switch bandwidth in the system, not HBM4 capacity per GPU. An earlier NVIDIA technical post said NVLink 6 provides 3,600 gigabytes per second of scale-up bandwidth per GPU.
NVIDIA says component specifications viewed in isolation do not determine application performance. A company representative said Vera Rubin performance depends on the co-design of models, software, computing, networking and memory. The representative said: "Individual component specifications do not determine application performance." At the system level, NVIDIA's page describes the Vera Rubin NVL72 platform as comprising 72 Rubin GPUs and 36 Vera CPUs. The company says the platform has entered full production and that systems based on it are being built and shipped.
The current figures follow earlier NVIDIA technical posts that presented higher targets for some connectivity and memory specifications. A post on the Rubin architecture said HBM4 delivers up to 22 terabytes per second, with 288 gigabytes of capacity per unit. In another post, the company said NVLink 6 provides 3.6 terabytes per second per GPU and 260 terabytes per second at the level of an NVL72 rack with 72 GPUs. NVIDIA also explained that its transition from Hopper to Blackwell doubled NVLink bandwidth and expanded scale-up from 8 GPUs to 72. The company linked this to a 50-fold improvement in MoE model inference performance per watt.
On supply, NVIDIA said in its disclosure for the second quarter of fiscal 2027 that production shipments of the Vera Rubin platform began in the third quarter of the fiscal year, with supply constraints continuing. The company said on its earnings call that its initial forecast for fiscal 2028 revenue growth was about 70% year over year, and that supply constraints could continue at least through the end of that fiscal year. It also expected Vera Rubin to account for about 20% of data center revenue in the third quarter of fiscal 2027. Its forecast for total revenue in that quarter was $108 billion, plus or minus 2%. The company had said shipment of the NVL72 rack was planned for the second half of 2026; the design includes 72 Rubin GPUs and 36 Vera CPUs distributed across 18 compute trays and 9 NVLink Switch trays.
In trading, NVIDIA shares closed at $213.66 on October 29, 2026, and then at $204.64 on October 30, according to EL7 data. These figures establish the change in closing prices between the two sessions without identifying its cause. The company had said third-party published benchmarks measured performance using production specifications. Its disclosure places the start of production shipments in the third quarter of fiscal 2027, while its forecasts point to supply constraints continuing at least through the end of fiscal 2028. The platform page says Vera Rubin NVL72 is in full production and shipping.