NVIDIA Rubin production starts in late 2025

The expert thinks NVIDIA still has a long time before Rubin can be made in large numbers.

Blackwell GPU & Vera Rubin, Source: NVIDIA/National Air and Space Museum

Vera Rubin is the patron of NVIDIA’s next-gen GPU architecture. Although NVIDIA has not officially disclosed this name yet, it has been circulating for a while following a leak from Kopite7kimi . The company has recently unveiled its most powerful GPU (AI Chip) to date, the Blackwell B100, set to launch later this year. However, with the AI craze in full swing, everyone is eagerly looking at NVIDIA’s roadmaps.

The R100 GPU is reportedly entering mass production by the end of 2025, while the system/rack solution is expected to start production no sooner than Q1 2026. It’s worth noting that these dates refer to production timelines and not shipping dates.

NVIDIA is still finalizing the specs and, more importantly, the size of the interposer and package. Currently, there are as many as three options under consideration, and it will be a 4x reticle design using CoWoS-L packaging, which is already utilized by Blackwell GPUs. Reportedly, Rubin would be employing the TSMC N3 process node (or perhaps that’s 3N).

• Nvidia’s next-generation AI chip, the R-series/R100 AI chip, will enter mass production in 4Q25, and the system/rack solution will likely start mass production in 1H26.
• R100 will use TSMC’s N3 node (vs. TSMC’s N4P for B100) and CoWoS-L packaging (same as B100).
• R100 adopts about 4x reticle design (vs. B100’s 3.3x).
• The interposer size for R100 has yet to be finalized. There are 2–3 options.
• R100 will be equipped with eight HBM4 units.
• GR200’s Grace CPU will use TSMC’s N3 process (vs. TSMC’s N5 for GH200 & GB200’s Grace CPU).
• Nvidia realizes that the power consumption of AI servers has become a challenge for customers’ procurement and data center construction. Hence, the R-series chips and system solutions focus on improving power consumption in addition to enhancing AI computing power.

— 郭明錤 (Ming-Chi Kuo)

One of the things that NVIDIA is said to be focusing on with the Rubin architecture is power consumption, as customers are facing challenges in providing enough power to their data centers. The Blackwell B200 GPU alone may consume up to 1000W.

Analysts reaffirm previous claims that Rubin is using 4x HBM4 stacks. Currently, there are no details on CPU/GPU superchips akin to GH200 or GB200.

Source: Medium