This is the second in a set of three blogs about projections for digital storage and memory for the following year that we have been doing for a while. Our first blog focused on the latest developments and projections for magnetic recording (HDDs and magnetic tape). 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, NVMe, NVMe-oF, including computational storage and CXL and how this will change the way we do computing.
At the close of 2023 demand for all storage and memory technologies has been down for about a year and a half for consumer, client and server applications. This was due to a large extent to the end of the pandemic and the reduction in consumer demand and the normalization of supply chains. In particular many storage customers and data centers were left with large inventories of storage products. As a consequence, NAND flash and HDD companies have cut back on production and generally down sized to control their costs.
The drop in NAND demand resulted in significant price declines through most of 2023 to the point where NAND flash was basically selling at the cost of production. As a consequence, in September Samsung cut its NAND flash production by 50% and other flash memory producers made similar production cuts. The NAND flash companies also reported that they focused their production on less than 128 layer flash.
TrendForce projected that this these cuts in unit and capacity production would stop the NAND flash price erosion in the 4th quarter of 2023 and could increase $/GB prices by 5%. The analyst firm recently said that enterprise SSD revenue grew 4.2% in 3Q23 and this was expected to increase by 20% in 4Q23. TrendForce also projected that DDR5 DRAM prices could rise 3-8% in 4Q23. WDC recently said that NAND flash prices could rise by 55% in the coming quarters. These are encouraging signs that the solid-state storage market is stabilizing and can look forward to solid growth in 2024.
There are many drivers for higher storage demand in 2024 and beyond and this should drive demand for storage devices. This looks like it will drive new capital spending going forward. Semi reported that memory-related capital expenditures saw a sharp decline in 2023. NAND equipment sales are predicted to have dropped by 49% to $8.8 billion in 2023 but it expects this spending will surge 21% to $10.7 billion in 2024 and rise another 51% to $16.2 billion in 2025. This in indicates projections for demand increases starting in 2024.
DRAM equipment sales are expected to remain stable, growing by 1% and 3% in 2023 and 2024, respectively. Supported by continuous technology migration and expanding demand for high-bandwidth memory (HBM), DRAM equipment segment sales are expected to increase an additional 20% to $15.5 billion in 2025.
NAND flash is the dominant primary storage (storage for active in-process data) in data centers and enterprise applications and it is often the only storage used in consumer devices such as smart phones and in most personal computers. In larger facilities active data lives on SSDs with colder data generally kept on hard disk drives (HDDs) and magnetic tape.
However, there are companies, such as Pure Storage, who project that with lower prices and better endurance using QLC NAND flash that it will displace demand for HDDs in secondary storage. Whether this happens to a significant degree in data centers depends upon whether HAMR HDDs are successful in enabling price declines for HDD storage, continuing the preference for HDDs for storing colder data.
At the 2023 Flash Memory Summit, SK hynix announced that they had made 321-layer 3D NAND flash (their 4D flash) but indicated that this would not go into mass production until 1H 2025. They said that this was possible by reducing the stress in the materials, with multiple overlay alignment corrections and with a more cost-effective process. This extended introduction date is in line with flash memory companies slowing their introduction of higher density memory until the market for NAND flash recovers.
SK hynix said that it started mass production of 238-layer NAND products in March 2023 for mobile devices. SK hynix also said that it was working on enhanced key-value computational storage devices for AI workloads. The company is also sampling processing in memory (PIM) products for AI applications.
Kioxia and Western Digital and other NAND flash vendors are moving to bonding separate logic and memory wafers together rather than making them on the same wafer. This allows a higher memory density and allows separate development of the logic and flash memory components. This approach was pioneers by Chinese flash memory company YMTC a few years ago.
NVMe is now the dominant flash memory interface and NVMe-oF (over fabric) where the fabric is often ethernet, is becoming common in data centers. NVMe-oF is being used to create pools of solid-state storage that can be shared between CPUs and servers.
This pooling and sharing of storage is referred to as disaggregation (breaking down the various parts of servers into a shared pool), which software can then use to create composable infrastructure supporting virtual devices or containers that can be created or destroyed as needed. This kind of pooling and composability is being extended to memory with the Compute Express Link (CXL) interconnect.
CXL provides for a switched network for memory of various types with different costs and performance. CXL 3.0 enables creating memory pools that can be shared between CPUs. There are some products being announced in 2023 for memory pooling with CXL and we expect some shipments of products including these advances in 2024.
Marvell, Microchip, Phision and other controller companies are supporting CXL (as well as NVMe) in their controllers as a means of achieving full data center composability that includes memory pooling as well as storage pooling.
There was much discussion this year about using NAND flash as part of upcoming CXL memory systems in 2023. This effort is partially driven by a demand for less expensive memory to support DRAM since the end of Optane memory development. SK hynix said that they are working on two types of CXL memory modules at the 2023 FMS, one focused on bandwidth memory expansion (BME) and one on capacity memory expansion (CME) as well as computational memory solutions (CMS). The slide below describes the BME, CME and CMS memory module devices.
Other major NAND companies, such as Samsung have also introduced their own CXL-based NAND flash products as well as controller company Microchip and independent SSD companies such as Fadu. Kioxia believes that higher CLX storage density using flash memory allows for less data transfers and increased processing efficiency for applications such as AI. They see using their BiCS flash for read intensive CXL applications and XL-Flash (with single level cells) for random read/write access memory expansion. Samsung is promoting its CXL semantic memory SSD for AI systems as shown below.
Kioxia also talked about putting more intelligence into their SSDs, with more powerful SoCs on enterprise SSDs that allow offloading RAID functions with their Gen 5 PCI products, to reduce CPU loads. They also talked about using flash memory storage for private AI models. Samsung spoke about SSDs for zonal control automotive applications. Samsung also announced high performance SSDs for generative AI applications. The company was showing its 256TB SSD that are geared towards building petabyte-scale solid state storage solutions.
Besides developments from the 2023 FMS there were also some relevant developments at the 2023 IEDM conference. For instance, Macronix has a method for healing wear using heat on NOR flash. They plan to use this on special areas on the die since the heating elements to do the refresh take up real estate. The company also had a paper on 3D NOR and 4Gbit die expected in 2024.
During WDC’s FMS talk looked into the various factors for increasing flash memory density and the impact they make on density improvements (see figure below). Vertical scaling (the number of layers) has a smaller contribution today than it did in the past since more layers means longer process time. Lateral scaling (the density of the memory holes can have the biggest impact, followed by logical density (bits per cell). Architectural changes, such as bonding logic and memory wafers together can also increase memory density.
Although Intel’s Optane memory started to wind down in 2022, various other non-volatile memory technologies are ramping up in embedded application, initially replacing NOR flash for code storage and some SRAM for cache memory. These include magnetic random-access memory (MRAM) and various resistive RAM (RRAM) technologies. TSMC, Samsung and other foundries have produced various embedded devices for wearable and automotive applications, in addition, the increasing popularity of chiplet technology and the new universal chiplet interconnect express (UCIe) interface may drive discrete memory chiplet demand, both for DRAM as well as emerging memories.
As the chart below from the Coughlin Associates and Objective Analysis Emerging Memories Branch Outreport indicates, growth in both embedded and discrete non-volatile memory technology (represented by MRAM) could drive increasing capacity shipments to a likely $59B of revenue by 2033.
At the 2023 IEEE MRAM Forum it appears that embedded memory applications are driving development and applications for emerging memories. In particular foundry companies such as Samsung are developing MRAM memory products to enable the next generation of automotive applications as shown below. This is driving development of 14nm and even down to 5nm products in the next few years.
Although 2023 ended with demand down for all types of storage and memory technologies for a second year, we expect that demand will recover in 2024 to meet increasing storage demand and to gain efficiencies from the latest technologies that include advances in NAND, DRAM, CXL and various emerging memories.
