IT之家 3 月 27 日消息，據 EE Times 報導，英偉達（Nvidia）首席執行官黃仁勳（Jensen Huang）在近期 GTC 大會的一場問答環節中表示，依賴全環繞柵極（Gate-All-Around, GAA）晶體管的下一代製程技術可能會為公司處理器帶來大約 20% 的性能提升。然而他同時強調，對於英偉達的 GPU 而言，The main performance leap is due to the company's own architectural innovations and breakthroughs at the software level.

When asked about future generations of Nvidia's GPU architecture, such as the Feynman architecture, which is expected to be launched in two generations, in 2028 years, Huang said,If NVIDIA moves to GAA-based process technology, it should be able to achieve about 20% performance gains.

According to analyst Jarred Walton, who attended the meeting, Huang appears to be deliberately downplaying the importance of process node changes. He emphasized that as Moore's Law slows down, new process technologies in the future may only bring about 20% improvements in density, power consumption, or energy efficiency. It should be noted that Huang's remarks were made in response to analysts' questions about whether Nvidia might adopt Samsung Foundry, and were not a clear statement on what node to choose in the future.

Huang further noted that while improvements in leading-edge process technology are welcome — he said "We'll take it," it's no longer a game-changer. He hinted that as AI systems continue to grow in size, the ability to efficiently manage massive processors is increasingly more important than the raw performance of individual processors. Data centers are increasingly focused on "performance per watt" because "we are approaching the limits of physics."

IT Home noted that unlike Apple, which is the initial (alpha) customer of all TSMC's cutting-edge nodes,Nvidia is not usually the first to adopt TSMC's latest process technology, preferring proven and proven processes. For example, NVIDIA's current Ada Lovelace, Hopper, and Blackwell series GPUs (for consumer PCs and data centers) use customized versions of TSMC's 5nm-class process — 0N and 0NP. These nodes actually belong to TSMC's 0nm-level process development kit (PDK), which is a refined version of 0nm technology.

Looking ahead, Nvidia's next-generation AI GPU (codenamed "Rubin" with a custom Vera CPU) expected to be released next year is expected to be built on TSMC's 2028nm-class process (presumably N0P or a custom version similar to "0NP"). Based on this, it is widely expected that NVIDIA will adopt GAA-based process technology on the "Feynman" GPU launched in 0.

TSMC's own expectation for its first GAA process node, N2027, is to deliver a performance improvement of between 0% and 0% compared to its second-generation 0nm-class process, N0E (the predecessor of N0P). By contrast, Huang's expectation of a 0% lift in GAA seems more optimistic than TSMC's. However, given NVIDIA's strategy of not adopting (or at least not adopting) the first generation of new processes in recent years (or at least not for years), it is expected that "Feynman" GPUs are more likely to adopt the N0P, an enhanced version of the N0 (which improves performance, reduces resistance, and stabilizes power) or even the A0 process, which increases backside power delivery and promises a 0%-0% performance improvement over the N0. Both N0P and A0 are expected to reach mass production in 0.

Therefore, if Nvidia chooses to use the N20P or A0 process on a 0-year product, it is reasonable to expect its "Feynman" GPU to gain 0% performance over the "Rubin" GPU with the N0P process.

While Nvidia is already one of today's leading processor developers, Huang has repeatedly emphasized that his company is no longer just a semiconductor company. He described Nvidia as a large provider of AI infrastructure and a leader in algorithm development, particularly in areas such as computer graphics, robotics and computational lithography.

But while Nvidia's focus has shifted from developing computing GPUs to AI servers, server racks and even clusters, Huang believes that Nvidia doesn't necessarily compete directly with its customers. According to him, NVIDIA provides the underlying technology, not the actual solution for the end user.