[ChipTalk] 'Preventing D-RAM Bending'... Leading 'High Bandwidth Memory' with Super Gap

High-Performance AI Memory Built by Stacking Multiple DRAMs
Significantly Increasing Data Transmission Paths with TSV Process

Samsung's TC-NCF, Hynix's MR-MUF
Focus on 'Hybrid Bonding' Eliminating Chip Space

The most notable keyword in this year's memory semiconductor market is 'High Bandwidth Memory (HBM)'. The popularity of HBM is expected to continue next year as well. Due to the AI effect, more places are seeking HBM, and forecasts suggest that demand will exceed supply through next year. This is why Samsung Electronics and SK Hynix, which boast over 90% market share in the HBM market, have high expectations.

HBM, a high-performance memory semiconductor for AI, stands for 'High Bandwidth Memory.' Here, bandwidth refers to the amount of data that can be transmitted within a certain period. The higher the bandwidth, the more data can be transmitted quickly. HBM has been able to increase processing speed by stacking multiple DRAM chips vertically and drilling tiny holes between the chips to dramatically increase the number of data transmission paths (I/O).

The technology used here is 'Through Silicon Via (TSV)'. TSV is a technique that drills tiny holes through a silicon wafer and fills them with conductive material to form electrodes (places where electricity flows). The electrodes created by drilling these holes serve as the I/O. While conventional technology has limitations in increasing the number of I/Os, the TSV process can significantly increase that number. In fact, the number of I/Os in HBM is 1,024, which is 32 times more than the 32 used in conventional DRAM.

Since its first development in 2013, HBM technology has advanced over 10 years under the leadership of the domestic memory industry. So far, first-generation (HBM), second-generation (HBM2), third-generation (HBM2E), and fourth-generation (HBM3) products have been released, and the next generation products will soon be commercialized. Samsung Electronics plans to mass-produce HBM3P, which adds 'Plus' to HBM3, while SK Hynix plans to mass-produce HBM3E, which adds 'Extended' to HBM3.

Notably, the two companies use different technologies to stack chips during the HBM production process. Samsung Electronics uses 'TC-NCF (Thermo Compression Non Conductive Film)', while SK Hynix uses 'MR-MUF (Mass Reflow Molded Underfill)'. TC-NCF refers to a method where a thin non-conductive film (NCF) is inserted between chips and then compressed with heat. When heated, the NCF film melts and acts as an adhesive, filling the space between the chips.

On the other hand, MR-MUF involves placing stacked chips into a space that functions like an oven and applying heat to perform a first soldering process. Then, a liquid protective material is injected between the chips to fill the space and solidified to complete the process. SK Hynix developed MR-MUF in collaboration with the Japanese materials company Namics. They received epoxy molding compound (EMC) used as a protective material from Namics, completed the technology, and have been using it since HBM2E.

Both companies are currently enhancing their respective technologies to improve the level of HBM products. Samsung Electronics is developing new NCF materials while advancing its independently developed NCF technology, applying it from HBM3 onward. They expect NCF technology to be useful in solving issues such as chip warpage that may occur as the number of stacked DRAM chips increases in future HBM production. They have also announced plans to introduce a new process called 'Hybrid Bonding,' which completely eliminates the space between chips and bonds them directly.

SK Hynix developed HBM3E by applying the latest MR-MUF technology, improving heat dissipation performance by 10% compared to the previous generation. They are currently sending samples to customers for performance verification and plan to begin full-scale mass production of HBM3E in the first half of next year. SK Hynix also aims to apply hybrid bonding in the future. However, since this process is technically challenging, they are developing it with a mid- to long-term plan.

This content was produced with the assistance of AI translation services.