Questa sezione non è ancora tradotta. Il contenuto seguente è mostrato in inglese, come pubblicato. Leggi la versione inglese.
NVIDIA lowers Rubin GPU memory and NVLink bandwidth in its shipping specs as first production racks arrive
What changed
ServeTheHome reports that NVIDIA's shipping specifications for the Rubin GPU list 19.2 TB/s of HBM4 memory bandwidth, down from a preliminary 22 TB/s, and 3.0 TB/s of NVLink, down from 3.6 TB/s, while scale-out networking rises from 0.4 TB/s to 0.45 TB/s. Compute (50 PFLOPS NVFP4 inference) and 288 GB of HBM4 are unchanged. The Vera Rubin NVL72 rack falls from 1,580 TB/s to 1,400 TB/s of memory bandwidth, and HGX Rubin NVL8 drops from 2.3 TB to 2.0 TB of HBM4. NVIDIA had labelled the earlier figures preliminary and has not said why they changed.
View original source (opens in a new tab)
Read the original: ServeTheHome: NVIDIA lowers Rubin GPU memory and NVLink bandwidth in its shipping specs as first production racks arrive (opens in a new tab)
We summarise in our own words and link the source. Figures are as reported by the source and not independently verified. Report a correction.
Related index data
A supplier link means the company is named in the story or offers this kind of product. It does not mean a contract unless the story says so.
Related on CoolPower Index: Chip power and cooling.
More in AI and Data Centre Chips
- CoreWeave's chief executive cancels his share sale plan after Q3 results, keeping 1.85 GW of active power as the end-2026 target 11 Nov 2026
- AMD makes 6th Gen EPYC 'Venice' server CPUs generally available, with Helion agentic AI racks in liquid-cooled and air-cooled versions 10 Nov 2026
- Anthropic's amended IPO filing raises its Microsoft compute commitment to at least US$47.2 billion and says a US$34.9 billion Lambda agreement can now be terminated 10 Nov 2026