How Far Have Samsung and SK Hynix Advanced in HBM?

②D-RAM 3 Companies Deliver HBM3E Samples to Customers
Samsung Electronics Uses 'Shinebolt'
SK Hynix Competes with HBM3E Name as Is

Samsung Electronics, SK Hynix, and Micron, the world's top three DRAM companies, have completed the development of the latest 5th generation HBM DRAM product, HBM3E, and have finished supplying samples to major customers including Nvidia. Depending on order situations, all three are expected to start mass production of HBM3E as early as the first half of next year, or at the latest, the second half. Along with the mass production of HBM3E, preparations will also begin for the development of the 6th generation HBM, HBM4.

Samsung Electronics recently unveiled HBM3E, named 'Shinebolt' in the United States, boasting high performance with a maximum data input/output pin speed of 9.8Gbps. This speed can process over 1.2TB (terabytes) of data per second (equivalent to processing 40 UHD movies of 30GB each in just one second). By optimizing NCF (Non-Conductive Film) technology, the chips are stacked tightly without gaps to achieve high-density stacking, and thermal conductivity has been maximized to improve heat characteristics.

Samsung Electronics is confident that it can take the lead starting from HBM3E, as the HBM market is still in its early stages. To relieve customers who order HBM from the hassle of separately outsourcing packaging to foundries like TSMC, Samsung offers a groundbreaking turnkey service combining the product with advanced packaging technology and foundry services tailored to the customers who order Shinebolt.

SK Hynix, which was the first in the world to enter the HBM market and has built trust through its technology, is emphasizing its expertise. Since HBM3E is an extended version of HBM3, SK Hynix is confident that it meets the world's highest standards in all aspects, including the speed essential for AI memory, heat control, and customer convenience, based on its exclusive mass production of HBM3 products.

SK Hynix's HBM3E can process over 1.15TB of data per second. The latest technology, MR-MUF (a process where liquid protective material is injected and solidified between stacked semiconductor chips to protect the circuits between chips), has been applied to this product, improving heat dissipation performance by 10% compared to previous products. HBM3E is also backward compatible, allowing customers to apply the new product without design or structural changes in systems configured with the previous HBM3.

Micron, a latecomer to the HBM market from the United States, plans to start mass production of HBM3E early in the 2024 fiscal year (September 2023 to August 2024). It aims to achieve annual HBM3E sales of about $700 million, which corresponds to approximately 2% of last year's revenue ($30.758 billion).

The reason why the world's top three DRAM companies are putting their lives on the line for HBM development and customer acquisition is that the HBM market is on the verge of explosive growth, and the price difference between HBM and general-purpose DRAM ranges from 6 to 7 times at the low end to over 10 times at the high end. Although HBM accounts for only 1% of the total memory market, the profit margin per product is much higher than that of general memory. For example, SK Hynix's HBM accounts for only 1% of total shipments but about 10-15% of sales and 40% of operating profit. The unit price is 10 to 15 times higher than conventional general-purpose DRAM, and the profit per unit sold is said to be 40 times that of existing products.

HBM orders are currently pouring in, mainly in the Americas. With companies like ChatGPT, Google, Meta, and Amazon entering the AI market, HBM, considered an essential item for AI, has become a product that cannot be sold fast enough. SK Hynix revealed in a conference call after its third-quarter earnings announcement that "all HBM to be produced next year is sold out." In other words, they have sold all the products that have not even been manufactured yet.

AMD stated that it is facing difficulties in procuring HBM required for its GPU product MI300X, scheduled for release at the end of the year, and is seeking breakthroughs by diversifying its existing supply strategies. Nvidia is also reportedly preparing to negotiate with major suppliers soon to secure additional HBM quantities.

As the main applications of memory expand from PC → mobile → data centers → hyperscale AI, the HBM market is expected to grow at a triple-digit annual rate. Market research firm TrendForce estimates that HBM supply will increase by 105% annually through next year, and semiconductor industry HBM sales will grow 127% to $8.9 billion next year compared to this year.

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