This is the third in a set of four blogs about projections for digital storage and memory for the following year that we have been doing for several years. Our first blog focused on the latest developments and projections for magnetic recording (HDDs and magnetic tape). The second blog addressed several optical storage startups.
This blog focuses on various types of solid-state non-volatile memory and storage. We will talk about the latest developments in flash memory, DRAM and non-volatile memories like MRAM, RRAM, FRAM and PCM as well as developing computer architectures using these technologies such as CXL and UCIe.
Memory and storage technologies began to recover in the 4th quarter of 2023 after a severe correction in 2022 and most of 2023. This correction impacted NAND flash and DRAM. However, 2024 has not shown steady growth , despite predictions that AI-powered PCs and smart phones would drive increased memory demand.
In November 2024 TrendForce indicated that the demand for current DRAM products such as DDR4 and LPDDR4X are trending downward due to declining demand and that advanced product demand for DDR5 and LPDDDR5X are uncertain going into 2025. As a consequence, more price declines by the end of Q4 2024 may occur. TrendForce has also revised its 2025 forecasts with the first half of 2025 having a pronounced drop in prices, particularly for the older DRAM technology.
However high bandwidth memory, HBM, will likely experience considerable growth in 2025 as NVIDIA and other suppliers of chiplet-based AI devices expand production to meet anticipated demand. For Q3 2024 the top four hyperscalers spent $58.9B on capex. This was a 63% annual growth rate and amounted to about 18% of their aggregate revenues of $293T. But it is not clear that the growth of revenues will match this capex spend, and if so, the hyperscalers could cut back on their capex spending perhaps in the second half of 2025.
Trendforce also said that the NAND flash market is also experiencing lower prices. This is especially true for NAND sold into consumer applications such as smart phones, where demand has tapered off.
On the enterprise and data center side the SSD companies are experiencing more demand and they are introducing new products such as high capacity, 60-100TB+, SSD with QLC flash to try and take some demand from nearline HDDs. At the 2024 FMS Samsung showed a chart, below projecting strong growth of high-capacity, depending upon the form factor, SSDs out to 2035.
SSDs are assuming primary storage for frequently used data in data centers and in particular for AI workflows and this is driving demand for these products. Active data lives on SSDs with colder data kept on hard disk drives, HDDs, and magnetic tape. The SSDs companies are also trying to introduce these high-capacity products as secondary storage for less active data, competing against nearline HDDs.
At the 2024 FMS Silicon Motion showed Gartner data showing that four bit per cell, QLC, flash is now more than 20% of the PC market and over 10% of the enterprise market. The image below is from Kioxia’s 2024 FMS keynote and shows the various places where SSDs are being used in AI processing.
However, as shown in the figure below, from Coughlin Associates, raw storage costs in $/GB for SSDs versus HDDs are still considerably higher and we expect this cost difference to be about 5X going into the future as it gets harder to pursue additional SSD cost reductions from increased layer count and as HAMR HDDs become mainstream, driving significant HDD capacity growth.
Recent NAND flash capacity increases have focused on smaller diameter memory cells, closer memory cells and particularly on using more bits per cell, rather than more NAND flash layers. At the 2024 FMS WDC and Kioxia talked about this being the end of the layers race as more layers increase the capital costs to make those layers. 2025 will likely introduce NAND flash products with more than 300 layers, but introduction of these products in volume has taken about 3 years.
The higher capacity SSDs being offered do have somewhat lower operating power and allow considerably more storage in a given rack space, compared to HDDs, which can be an advantage in some applications. However, the QLC flash has lower endurance and thus is best used in write seldom applications and probably will have a limited impact on the demand for nearline storage using HDDs.
The market for non-volatile memories such as magnetic random access memory, MRAM, resistive RAM, ReRAM, ferroelectric RAM, FRAM, and phase change memory, PCM, continues to develop, particularly for embedded memories. In our 2024 report on non-volatile memories, A Deep Look at New Memories, we looked into developments and projections.
The first widespread use of emerging memories is for embedded memory on a CMOS logic chip, replacing NOR flash, which has reached its scaling limit at 28nm, and is today frequently being replaced with MRAM and ReRAM. Single-transistor MRAM cells are now competing with multi-transistor static random access memory, SRAM, to dramatically reduce the number of memory transistors on a chip in order to provide a lower-cost, higher-density solution. Several enterprise, industrial and consumer devices currently use MRAM as an embedded memory, and this trend will continue.
At the 2024 IEEE MRAM Forum, at the end of the 2024 IEDM conference in San Francisco and organized by the IEEE Magnetics Society and the IEEE Electron Devices Society, Kevin Garello, from CEA in France, showed a slide summarizing some of the stand-alone and embedded non-volatile memory applications.
TSMC is a leader in introducing MRAM into embedded devices and its MRAM development plan down to 5nm devices is shown below. This includes NOR flash like MRAM or eMRAM and RAM-like MRAM with higher endurance and performance. These devices are tailored for various automotive and other embedded applications. TSMC makes embedded chips for many other companies and its MRAM are being used in many of these devices. TSMC also offers ReRAM non-volatile memory as well as MRAM.
Based upon considerations of economy of scale, that is, the ability to drive up production volume over time, the report projects that total baseline emerging memory annual shipping capacity will rise from an estimated 340TB in 2023 to 8.46EB in 2034. Total emerging memory baseline revenues are expected to increase from $421M in 2023 to about $71.7B by 2034. The bulk of this rapid revenue growth will be supported by emerging memories’ displacement of SRAM, NOR flash and some DRAM.
AI is driving demand for DRAM and NAND although the big growth is in data centers, with some concern about the current rate of data center capex spending. Products using non-volatile memories in embedded applications are growing.
