Nvidia CEO Jensen Huang, a leading player in Artificial Intelligence (AI) chips, has officially announced the adoption of the next-generation AI semiconductor chip, Rubin, and the sixth-generation high-bandwidth memory (HBM), HBM4. The release cycle of the AI semiconductor chip is also expected to heat the HBM development speed competition by bringing it forward from the existing two years to one year.
Huang partially revealed the detailed specifications of Rubin, which is set to be unveiled in 2026, ahead of Computex 2024 in Taipei, Taiwan, on June 4. Rubin is the next version of Blackwell, which was introduced this year. Rubin is named after Vera Rubin, an American astronomer who studied dark matter in space and the rotation speed of galaxies. Nvidia also plans to launch its own Central Processing Unit (CPU) called Vera soon.
So far, Nvidia has introduced new architecture every two years. In 2020, the A100 based on Ampere was released, and in 2022, the Hopper-based H100 was launched. The H100 is the most favored GPU for AI accelerator purposes by global big tech companies investing in AI infrastructure. In March, the Blackwell-based B100 was unveiled. The B100 is scheduled to start mass production in the third quarter and start shipping at the end of the year. However, Nvidia has announced a new model even before the previous model enters mass production. Huang said, “GPU development will proceed every year after Rubin.”
Accordingly, Nvidia will release Blackwell Ultra next year, Rubin in 2026, and Rubin Ultra in 2027. Blackwell Ultra dramatically improves the performance of the server GPU that Nvidia released in March 2024. Rubin has a completely redesigned internal structure.
Specifically, Rubin plans to implement an AI semiconductor chip by applying HBM4 to the Graphics Processing Unit (GPU). This is the first time Nvidia has revealed whether HBM4 will be mounted on the next-generation semiconductor chip. The HBM currently developed by SK Hynix, Samsung Electronics, and Micron is the 5th generation HBM3E. HBM4 is a next-generation product currently under development by memory companies and is expected to be completed by 2025.
Rubin will contain 8 HBM4s and Rubin Ultra will contain 12 HBM4s. The demand for HBM is expected to increase. Accordingly, whether SK Hynix and Samsung Electronics will benefit is attracting attention. Unlike Blackwell, which is mass-produced in a 4-nanometer process, Rubin uses a 3-nanometer process.
Meanwhile, Nvidia has unveiled a new product lineup of RTX AI PCs equipped with GeForce RTX GPUs. The newly introduced RTX AI PCs include the Asus TUF A14/A16, Zephyrus G16, ProArt PX13/P16, MSI Stealth A16 AI+, and more.
Nvidia explained that content creators can simplify and automate their workflow using the RTX-based AI PC, while streamers can use AI-based background and noise removal features. In addition, they conveyed that by utilizing the RTX AI Toolkit, which helps developers optimize and personalize their work, tasks can be processed quickly and with less capacity.
