Samsung tipped to supply HBM3 to Nvidia for Chinese market

Samsung Electronics Co. has passed Nvidia Corp.’s qualification tests for its fourth-generation high-bandwidth memory, or HBM3 chips, for use in its AI application processors for the first time, Reuters reported on Tuesday. The report raised market expectations that Samsung’s fifth-generation HBM, or HBM3E, would also be cleared by the AI chipset maker soon.

Samsung's HBM3 chips will, for now, only be used in a less sophisticated Nvidia graphics processing unit (GPU), the H20, tailored for the Chinese market, according to the report citing unidentified multiple sources. The South Korean chipmaker could begin supplying the HBM3 to Nvidia as early as August.

It was not immediately clear if Nvidia would use Samsung's HBM3 chips in its other AI processors, or if the chips would have to pass additional tests before that could happen, Reuters added.

The H20 is the most advanced of three GPUs Nvidia has tailored for the Chinese market after the US tightened export restrictions in 2023.

HBM3E

The report relieved investor doubts about whether Samsung’s HBM3E, the most advanced memory in the high-value, high-performance chip segment, could pass Nvidia’s quality tests.

In May, Reuters reported Samsung’s HBM3E has yet to meet Nvidia's standards. Samsung denied the report as untrue.

The world’s No. 1 memory chipmaker is striving to catch up with local rival SK Hynix Inc. the unrivaled leader in the market of HBM chips, essential for processing a vast amount of data for generative AI applications.

Samsung began mass production of HBM3 chips in the third quarter of last year.

But SK Hynix stays ahead in the race. This year, the world’s No. 2 memory chip producer became the industry’s first to mass-produce the HBM3E, for shipment to Nvidia.

SK Hynix controls 53% of the HBM market, trailed by Samsung with a 38% share. Micron Technology Inc. claims the remainder of the market.

TECHNOLOGY WAR

HBM is made of vertically interconnected multiple dynamic random access memory chips. It saves space and reduces power consumption.

The HBM3E chip SK Hynix began supplying to Nvidia in March this year is composed of eight layers of DRAMs. Micron also launched the mass production of 8-layer HBM3E in February this year for supply to Nvidia.

In February of this year, Samsung unveiled a more advanced HBM: the industry’s first 12-layer HBM3E that boasts the largest memory capacity of 36 GB for HBM.

But it has not yet passed Nvidia’s qualification tests.

Samsung was reportedly seeking to verify with Nvidia the qualifications of its 8-layer and 12-layer HBM3E chips by the end of the third quarter and fourth quarter of this year, respectively.

Write to Jeong-Soo Hwang at hjs@hankyung.com
 


Yeonhee Kim edited this article.