What is the Real Reason for NVIDIA to Need Samsung Electronics? – BusinessKorea
In March this year, NVIDIA CEO Jensen Huang unveiled the B100, based on the more powerful 'Blackwell' architecture, at the GTC 2024 held in San Jose, USA, drawing significant attention. CEO Huang announced that production of the B100 would begin in the third quarter of this year.
During the keynote speech at Computex 2024 held in Taipei, Taiwan, on June 2, CEO Huang introduced the next-next-generation AI semiconductor 'Rubin' for the first time. The Rubin-based R100 will include eight sixth-generation HBM (HBM4) units, while products based on the more powerful Rubin Ultra will feature twelve HBM4 units.
On June 4, at a press conference in Taipei, CEO Huang stated, "We are also working hard to ensure that they (Samsung Electronics, Micron) pass the tests as quickly as possible so that they can be used in our AI semiconductor processes." He emphasized the importance of high-bandwidth memory (HBM) in their AI accelerator lineups, which include the H100, H200, B100, and B200. "The memory speed required for these is significant, making HBM very important," Huang added.
According to industry sources on June 29, NVIDIA will launch the B100 in the third quarter. The B100 will need eight HBM3E units. Currently, SK Hynix is virtually monopolizing the supply of HBM3 and HBM3E to NVIDIA, but industry sources explain that it is insufficient to meet NVIDIA's needs. Especially with the B100 release in the third quarter, the demand for HBM from NVIDIA is expected to continue increasing, leading to anticipated supply shortages.
Park Jun-young, a researcher at Hyundai Motor Securities, stated, "Without the help of the other two companies besides SK Hynix, NVIDIA will not be able to meet its HBM demand." Park predicted, "The HBM shortage situation will continue until the fourth quarter of this year, and if the three companies' CAPA is not included in the global HBM CAPA, the shortage could extend to the first half of 2025."
CEO Huang has set the mass production timeline for Rubin-based products for 2026. Samsung Electronics and SK Hynix have announced their goal to develop and mass-produce HBM4 by next year. NVIDIA has received HBM3E samples from Samsung Electronics for testing, but no supply contract has been announced yet. Micron's HBM production capacity lags behind its competitors, and yield issues complicate resolving the current HBM shortage.
The current flagship AI semiconductor from NVIDIA, the Hopper-based H100, contains five fourth-generation HBM (HBM3) units, while the upcoming H200 will feature six fifth-generation HBM (HBM3E) units. The B100 will need eight HBM3E units, highlighting the increasing demand for high-bandwidth memory.
As the tech industry closely watches these developments, the collaboration between NVIDIA and its HBM suppliers will be crucial in determining the success of its next-generation AI semiconductors. The industry anticipates that the HBM shortage could have significant implications for production, pricing, and competition within the tech sector.
Never miss another article! Enter your email address below to receive our free daily newsletter:
