HBM5 Thermal Showdown: Samsung HPB vs. SK Hynix iHBM, Micron Accelerates Catch-Up

Source Tradingkey

TradingKey - Samsung Electronics debuted its eighth-generation High Bandwidth Memory HBM5 prototype at Computex 2026 on Tuesday and announced that it has delivered 12-layer HBM4E samples to customers. SK Hynix, on the eve of the COMPUTEX conference, unveiled its differentiated heat dissipation technology 'iHBM,' with mass production of HBM5 expected between 2029 and 2030. Micron Technology ( MU) is also accelerating its catch-up efforts. This competition is shifting from a race for speed to a battle over heat dissipation and packaging capabilities, and NVIDIA ( NVDA ), as the largest customer, its choice is crucial.

Samsung Electronics: HPB "Chimney" Cooling to Enter Mass Production in 2028

Samsung showcased a physical HBM5 prototype at its booth. Song Jai-hyuk, the company's CTO, told Reuters: "As AI systems become increasingly complex, full-stack capabilities integrating memory, foundry, and packaging will be the deciding factor."

HBM5 will use Samsung's 2nm process for the base die and introduce HPB thermal management technology—a copper-based heat-conducting structure embedded in the chip that creates independent heat transfer channels to significantly reduce thermal resistance, which he vividly referred to as a "chimney."

Samsung plans for HBM5 to offer 12-layer, 16-layer, and 20-layer stacks, aiming for mass production around 2028.

SK Hynix: First to Unveil iHBM, Reducing Thermal Resistance by 30%

On May 26, just one week before the opening of Computex, SK Hynix preemptively unveiled its "iHBM" temperature control and heat dissipation technology, specifying its use for HBM5.

The technology embeds an integrated cooling element named "ICE" directly into the D2D PHY region—where heat concentration is highest—to build a dedicated heat dissipation channel, reducing thermal resistance by more than 30% compared to traditional solutions while continuing to use the mature MR-MUF packaging process to ensure compatibility with customers' existing systems.

Lee Kang-wook, Vice President and Head of Packaging Development at SK Hynix, stated: "iHBM combines memory design with advanced packaging technology and is the optimal solution for achieving minimum heat generation." However, industry analysis suggests that SK Hynix’s mass production timeline for HBM5 is set for around 2029 to 2030, approximately one year later than Samsung.

SK Group Chairman Chey Tae-won stated during the event that the company plans to double its wafer capacity over the next five years, reiterating that "the shortage of memory chips driven by AI will persist until 2030."

Micron accelerates catch-up, showcases full AI storage product lineup.

Micron Technology showcased a comprehensive portfolio of AI-optimized memory and storage solutions at Computex, covering multiple product lines including HBM4, LPDDR, DDR, GDDR, and enterprise-grade SSDs.

Sumit Sadana, the company's executive vice president, stated: "System performance is now increasingly driven by memory bandwidth and memory capacity. This structural shift in the semiconductor ecosystem has made memory and storage indispensable strategic assets."

Notably, Micron has announced that its entire HBM capacity for 2026 is fully sold out and has raised its HBM market size projection to reach $100 billion by 2028, two years ahead of its previous forecast.

At the GTC conference in March 2026, Micron also confirmed that it has begun mass shipping HBM4 36GB 12-layer stacked products designed for NVIDIA's Vera Rubin platform.

As of now, NVIDIA, the largest buyer of HBM, has not publicly commented on the two thermal solutions from Samsung (HPB) and SK Hynix (iHBM). According to sources, NVIDIA is currently validating technologies from both companies, and the final choice will depend on mass-production yields and system integration efficiency.

Overall, heat dissipation and packaging yield have become the core bottlenecks in HBM competition. Whoever can achieve the first breakthroughs in stacking layers and thermal management will win the next generation of AI orders.

Disclaimer: For information purposes only. Past performance is not indicative of future results.
goTop
quote