At last week's GTC conference, NVIDIA further clarified the roadmap for the next-generation GPU architecture Feynman, which is expected to be released in 2028. Feynman will not only use custom HBM chips, but will also use Die Stack, a 3D stack packaging process, for the first time.Among them, the GPU core will launch TSMC’s 1.6nm process-A16.

This is also NVIDIA’s first launch of TSMC’s new process in many years. The last launch was in the 55nm era. After the rise of the mobile era, Apple almost dominated the launch of TSMC’s new process. This time it has happened again with the rise of the AI ​​era.

The A16 process is an improved version of the 2nm process N2. It will add an additional new technology to the GAA transistor - SRP back power supply, which improves density and performance, and also improves the power supply capacity. It is therefore suitable for HPC computing, but it is not a low-power chip, and Apple will not be the first to launch it.

According to TSMC, the A16 process can improve performance by 8-10% compared to the 2nn enhanced N2P process, or reduce power consumption by 15-20%, and increase transistor density by 7-10%.

Although NVIDIA has obtained the A16 process for the first time, it faces many challenges.The main reason is that production capacity is limited and the improvement is not as expected.A16's production capacity will only be 20,000 wafers per month by the end of next year, and will double to 40,000 wafers in 2028. The entire 2nm process family's production capacity is expected to reach 200,000 wafers per month, so A16's production capacity ratio is not high.

This also resulted in the Feynman GPU having to make changes, and could only use the A16 process on the core GPU Die, which is the most sensitive to performance and power consumption.Less core parts will use TSMC’s N3P process, which will also help reduce costs.

In fact, NVIDIA has long realized that relying solely on TSMC for advanced production capacity would be too risky and not conducive to price negotiations, so it is also actively seeking other foundries. The LPU chips released this time are manufactured by Samsung, and it is possible to use Intel's EMIB-T technology to package chips in the future. If the cooperation goes smoothly, it is not impossible to use Intel's 14A process to manufacture GPU chips.