Samsung set to triple HBM output in 2024 to lead AI chip era

SILICON VALLEY – Samsung Electronics Co., the world’s largest memory chipmaker, will likely triple its high bandwidth memory (HBM) chip production volume this year from last year’s output as it aims to take the lead in the artificial intelligence chip segment.

At Memcon 2024, a gathering of global chipmakers, held in San Jose, California on Tuesday, Hwang Sang-joong, corporate executive vice president and head of DRAM Product and Technology at Samsung, said he expects the company to increase its HBM chip production volume by 2.9 times this year compared to last year’s output.

The forecast is higher than Samsung’s projection unveiled at CES 2024 early this year that the chipmaker will likely produce 2.5 times more HBM chips in 2024.

“Following the third-generation HBM2E and fourth-generation HBM3, which are already in mass production, we plan to produce the 12-layer fifth-generation HBM and 32 gigabit-based 128 GB DDR5 products in large quantities in the first half of the year,” Hwang said at Memcon 2024.

“With these products, we expect to enhance our presence in high-performance, high-capacity memory in the AI era.”

At the conference, Samsung unveiled its HMB roadmap, which envisions 13.8 times more HBM shipments in 2026 than the 2023 output. Annual HBM output volumes will rise further to 23.1 times the 2023 level by 2028, it said.

In HBM4, the sixth-generation HBM chip code-named “Snowbolt,” Samsung plans to apply the buffer die, a control device, to the bottom layer of stacked memory for efficiency.

At the conference, Samsung allowed participants to demonstrate its latest HBM3E 12H chip – the industry’s first 12-stack HBM3E DRAM, marking a breakthrough with the highest capacity ever achieved in HBM technology.

Samsung is currently sampling its HBM3E 12H chips to customers and plans to start mass production in the first half.

The conference participants included SK Hynix Inc., Microsoft, Meta Platforms, Nvidia and AMD.

CXL TECHNOLOGY AT MEMCON 2024

Samsung also unveiled at Memcon 2024 the expansion of its Compute Express Link (CXL) memory module portfolio, showcasing its technology in high-performance and high-capacity solutions for AI applications.

In a keynote address, Choi Jin-hyeok, corporate executive vice president of Samsung’s Device Solutions Research America, said Samsung is committed to collaborating with its partners to unlock the full potential of the AI era.

“AI innovation cannot continue without memory technology innovation. As the memory market leader, Samsung is proud to continue advancing innovation, from the industry’s most advanced CMM-B technology to powerful memory solutions like HBM3E for high-performance computing and demanding AI applications.”

Highlighting growing momentum in the CXL ecosystem, Samsung introduced its CXL Memory Module – Box (CMM-B), a cutting-edge CXL DRAM memory product.

The chipmaker also showcased CXL Memory Module Hybrid for Tiered Memory (CMM-H TM) and CXL Memory Module – DRAM (CMM-D) technology.

CXL is a next-generation interface that adds efficiency to accelerators, DRAM and storage devices used with central processing units (CPUs) in high-performance server systems.

LAGGARD IN HBM SEGMENT

A laggard in the HBM chip segment, Samsung has heavily invested in HBM to rival SK Hynix and other memory players.

HBM has become essential to the AI boom as it provides faster processing speed than traditional memory chips.

Last week, Kyung Kye-hyun, head of Samsung's semiconductor business, said the company is developing Mach-1, its next-generation AI chip, with which it aims to upend its crosstown rival SK Hynix, the dominant player in the advanced HBM segment.

The company said it has already agreed to supply Mach-1 chips to Naver Corp. by the end of this year in a deal worth up to 1 trillion won ($752 million).

With the contract, Naver hopes to significantly reduce its AI chip reliance on Nvidia, the world’s top AI chip designer.

Write to Jin-Suk Choi at iskra@hankyung.com

In-Soo Nam edited this article.