This type of storage will replace DRAM

Persistent memory has been expected to bring about a paradigm shift in computing, but that's not likely to happen anytime soon. In a recent webinar, industry insiders from the Storage Networking Industry Association (SNIA) expressed confidence that new technologies will replace existing memory technologies such as DRAM, but may not appear before the end of the decade.

In response to questions from media and industry observers, SINA's Arthur Sainio, Tom Coughlin and Jim Handy said that persistent memory has matched the speed achieved by modern DRAM technology, This is demonstrated by the hafnium ferroelectrics of SK Hynix and Micron. However, they cannot directly answer which emerging memory technologies will eventually replace DRAM in client PCS and servers.

Although ferroelectric memory is known for its fast write cycle, there is no guarantee that it will eventually prevail. This is because multiple new memory technologies such as MRAM, FERAM, and ReRAM are racing to replace existing standards such as SRAM, NOR flash, and DRAM.

MRAM has a significant advantage over its competitors as its read speed "will likely be comparable" to DRAM speed in the near future. New technologies such as spin-orbit torque and voltage-controlled magnetic anisotropy have also reduced the write latency of MRAM, making it one of the leading candidates that could one day replace DRAM.

However, a big hurdle in transitioning from DRAM to persistent memory is manufacturing cost. While DRAM is relatively inexpensive to produce, persistent memory can take years to become competitive in terms of price.

Another issue hindering persistent memory adoption is that it currently uses NOR flash and SRAM interfaces instead of DDR. However, this may change in the future, as "no memory technology has the inherent characteristics of being tightly coupled to any type of bus."

Persistent memory, as the name suggests, can retain content even when there is no power supply, which makes it a huge asset in some applications. However, we believe that despite its obvious benefits, there are still many obstacles to its widespread adoption in the near future. As things stand, we may not transition to new technologies before the early 2030s, "but it could be much later."

Next generation DRAM, What to focus on?

During the "winter" of the memory market, demand for DRAM bits remained weak, except for some specific applications such as AI servers and automotive electronics. In particular, the rise of generative AI applications such as ChatGPT since the end of 2022 has driven the demand in the data center market for high-speed memory technologies, namely DDR5 DRAM and High bandwidth memory (HBM).

To capitalize on the new wave of generative AI and accelerate the market recovery, Samsung, SK Hynix, and Micron began shifting more wafer capacity to capture the HBM opportunity, slowing overall bit production and accelerating the shift to under-supply for non-HBM products. Total HBM wafer production is expected to increase by more than 100% year-over-year to approximately 150 kwpm by 2024.

Driven by strong demand for AI computing, HBM growth will largely outpace the overall DRAM market. After an impressive growth in bit shipments in 2023 (+ 93% YoY), HBM Bits are expected to continue to grow strongly in 2024 (+ 147% YOY) and over the next five years (CAGR of approximately 45%23-29). By comparison, CAGR23-29 for data center DRAM bits is about 25%.

In terms of revenue, the HBM market has the potential to grow from about $2.7 billion in 2022 to about $14 billion in 2024, which equates to about 3% and about 19% of overall DRAM revenue, respectively.

DRAM technology for the future

4F2 DRAM is currently being developed, which reduces the chip area by approximately 30% compared to the existing 6F2 architecture and eliminates the need for smaller lithographic nodes. In addition, after the introduction of the 4F2 unit from 2027, the CMOS bonded array (CBA) DRAM architecture will be adopted, where the peripheral circuits and memory arrays are machined on different wafers and then stacked together by interwafer hybrid bonding.

HBM also requires hybrid bonding to continue to improve memory bandwidth and power efficiency, and to minimize HBM stack thickness. We expect hybrid bonding to be adopted starting with HBM4 generation (~2026), which will employ 16 DRAM chips per stack and double the interface width to up to 2,048 bits.

While it is still finding its way, single-chip 3D DRAM is an important long-term scaling solution. Many key aspects of 3D-DRAM development remain unclear, and the strategy moving forward has yet to be determined. According to technical papers and applications, DRAM companies are exploring a number of possible paths to achieve monolithic 3D-DRAM, including 1T-1C cells with horizontal capacitors, as well as capacitor-less options such as gain cells (2T0C) or 1T-DRAM (for example, based on floating body effects). The transition from 2D to 3D DRAM will trigger a major shake-up in the DRAM industry, essentially mimicking the history of 3D NAND. Our current market model assumes 3D DRAM will hit the market around 2030 and will take about five years to reach 10 mwpy (38% of forecast DRAM wafer production in 2035).

The battle for HBM's leadership is set to escalate

SK Hynix has been a pioneer in the development and commercialization of HBM since 2013 and currently leads the HBM market with approximately 55% revenue share, followed by Samsung with approximately 41%. Micron is notably absent from HBM until 2020. As a latecomer to the HBM business, the US company needs to shorten its time to market, and to do so, they will skip the HBM3 generation and go straight to the HBM3E product (aka HBM3 Gen2).

Advanced packaging technologies for DRAM logic integration and DRAM miniaturization replacing traditional lithography miniaturization technologies will be the focus of R&D activities for Chinese companies in the coming years. These paths are currently open - there are no significant obstacles due to commercial restrictions at this stage - and China will follow them to develop high-performance AI chips without using restricted cutting-edge equipment. By leveraging hypermol solutions that combine logic and storage, China will continue to compete for technological supremacy in various fields, key among which is AI computing.