Short: CEO Jensen Huang briefly touched the company’s GPU Roadmap beyond the upcoming Rubin Architecture, covering a wide range of subjects during its keynote speaker at NVDia’s GTC. Fenman generations and those who follow will probably adopt semiconductor nodes with gate-all-round (GAA) transistors.
During a Q&A session at the GPU Technology Conference, Huang estimated the uplift of about 20% of the performance in the GAA architecture. Their comments provide an initial signal of the company’s expectations for future graphics chip design.
Semcover manufacturers such as TSMC and Intel have postponed the GAA transistor as a major technique for process nodes beyond 3Nm. Recent nodes, including 3Nm, have employed the finfate, a structure where the transistor gate surrounds the current channel on three sides. However, the bleeding-dive transistors shrink and pack more densely in advanced nodes, the electric leakage becomes a growing concern.
See also: How CPU is made, design process – Touching the basics of how the transistor works
The GAA addresses the issue by vertically stacking the current channels, which increases the total channel area, while the gates allow the gates to surround the channels on all four sides. Although technology is more expensive, it increases both performance and energy efficiency.
Intel’s upcoming 18A node will introduce GAA with Panther Lake Laptop CPU and Clearwater Forest Server Processor later this year. TSMC plans to adopt GAA for its 2Nm N2 process, which is close to production and is expected to debut in the A20 SoC of the iPhone 18 Pro at the end of 2026.
In GTC, Huang officially revealed the next GPU generation of NVIDIA, Vera Rubin, which will be built on the N3 node of TSMC and expected to be available to the enterprise customers next year. He also shared initial details about Rubin’s successor, Fenman, which is estimated to come in 2028.
Although information about the semiconductor process of Fenman is coming upcoming, it is likely to use 2Nm-class node, although NVIDIA can also consider Intel’s 18A. Either in the case, GAA will be a major component. Huang commented that it is not revolutionary to promote 20% of the performance from GAA, “whatever we can get.”
Meanwhile, backside power delivery is emerging as another important technique for sub -3Nm nodes. TSMC is not expected to implement it until the following node A16 starts from 2026, while Intel will introduce its version with 18A later this year.
Interestingly, during GTC Q&A, Huang also described NVIDIA as “AI Infrastructure Company”. While Nvidia increased prominently for its gaming GPU, AI Boom has inspired the value of its enterprise division into the multi-trilian-dollars range.
