In short: Whereas masking a variety of subjects throughout his keynote at Nvidia’s GTC, CEO Jensen Huang briefly touched on the corporate’s GPU roadmap past the upcoming Rubin structure. The Feynman technology and people who comply with will seemingly undertake semiconductor nodes that includes gate-all-around (GAA) transistors – the subsequent evolution in transistor know-how after FinFET.
Throughout a Q&A session on the GPU Expertise Convention, Huang estimated a roughly 20% efficiency uplift from transitioning to GAA structure. His feedback supply an early indication of the corporate’s expectations for future graphics chip designs.
Semiconductor producers resembling TSMC and Intel have touted GAA transistors as a key know-how for course of nodes past 3nm. Current nodes, together with 3nm, have employed FinFET, a construction the place the transistor gate surrounds the present channel on three sides. Nonetheless, as bleeding-edge transistors shrink and turn into extra densely packed in superior nodes, electrical leakage turns into a rising concern.
Additionally see: How CPUs are constructed, the design course of – touches on the fundamentals of how transistors work
GAA addresses this challenge by stacking present channels vertically, growing the entire channel space whereas permitting gates to encompass the channels on all 4 sides. Though the approach is dearer, it enhances each efficiency and vitality effectivity.
Intel’s upcoming 18A node will introduce GAA with Panther Lake laptop computer CPUs and Clearwater Forest server processors later this 12 months. TSMC plans to undertake GAA for its 2nm N2 course of, which is nearing manufacturing and is anticipated to debut within the iPhone 18 Professional’s A20 SoC in late 2026.
At GTC, Huang formally revealed Nvidia’s subsequent GPU technology, Vera Rubin, which will probably be constructed on TSMC’s N3 node and is anticipated to turn into out there to enterprise shoppers subsequent 12 months. He additionally shared early particulars about Rubin’s successor, Feynman, which is anticipated to reach in 2028.
Though data on Feynman’s semiconductor course of is forthcoming, it is more likely to make the most of a 2nm-class node, although Nvidia may think about Intel’s 18A. In both case, GAA will probably be a key element. Huang remarked that whereas a 20% efficiency increase from GAA is not revolutionary, “we’ll take what we are able to get.”
In the meantime, bottom energy supply is rising as one other vital know-how for sub-3nm nodes. TSMC will not be anticipated to implement it till the next node known as A16 starting 2026, whereas Intel will introduce its model with 18A later this 12 months.
Curiously, in the course of the GTC Q&A, Huang additionally described Nvidia as an “AI infrastructure firm.” Whereas Nvidia rose to prominence for its gaming GPUs, the AI growth has propelled the worth of its enterprise division into the multi-trillion-dollar vary.
