When most designers hear the word “memory” in the context of electronics, they probably think of Flash or DDR3/4. These technologies are certainly popular, but other emerging technologies are making headway for application-specific embedded systems. Even as the DDR5 specification is rolling off the presses, legacy memories will still have their place in certain embedded applications.
For embedded designers, there is a wealth of memory options available for new systems. New products are being released across memory types, even as the larger memory vendors focus on bigger customers with smaller part number counts. Ironically, some of these new emerging memory products are not new at all as older memory types still have a role to play in new products.
Matching Legacy and Emerging Memory to Applications
Memory is one of those components that won’t go away, even though a more advanced type of memory hits the market. As larger companies like Samsung have labeled their earlier DDR products EOL and focused on the latest-and-greatest, smaller companies have stepped up with broad product portfolios that include everything from NAND flash to DDR2. Embedded systems designers can still access these earlier products, either as standalone high-capacity chips or integrated into processors.
Even as leading semiconductor companies focus on the newest iterations of proven technology (e.g., DDR5 and soon DDR6), experimental types of volatile and non-volatile memory are the subject of intense research and commercialization. The goal is to develop products that can support upcoming technologies like AI, edge computing, self-driving cars, and other devices we may not have conceived. The table below shows some of these application areas for older and newer memories for comparison.
Emerging Memory for Embedded Products
With the range of memory applications being as diverse as product lines on the market, it’s doubtful any one of these will replace DDR4 and higher for general-purpose computing. Instead, given the unique features of upcoming RAM types, they’ll likely be confined to some niche applications in embedded systems, data centers, mobile devices, intelligent systems, and many other areas. Let’s look at some of these emerging memory types.
FRAM and MRAM
These two technologies are worth comparing as they are both magnetic, but MRAM is decidedly more advanced and targeted at more advanced applications. It is true that non-volatile FRAM modules are available on the market, but they are stalled at 4-8 MB. FRAM read/write cycles are also destructive with low latency (~50 ns), so these modules are not applicable for high-speed, high-capacity systems. Some application areas include:
Fujitsu 1 MB FRAM module. From the Fujitsu MB85RS1MTPNF-G-JNERE1 datasheet.
MRAM adoption is low, but only because it has not been on the market very long and foundries are still investing in production capacity to satisfy projected demand. MRAM stores each bit of data as a magnetic orientation, and applying a voltage gives an MRAM device some probability of changing state. This is actually useful in an application like neural networking, where random weight initialization could be used for embedded AI systems. This technology may be useful in low-power AI ASICs and SoC, particularly within the AI compute block.
Currently, 40 nm ReRAM has achieved technical qualification for use in consumer products, and 22 nm ReRAM has been in risk production since 2019. Bringing high-density ReRAM into practical applications requires overcoming a number of technical and manufacturing challenges, so we shouldn’t expect the next round of laptops to run on ReRAM.
The current ideal application for low-density ReRAM memory arrays is for parallel neural network processing in conventional computing. With ReRAM’s nearest competitor being Flash, ReRAM offers faster read/write latency and lower power consumption, making it useful in embedded systems requiring fast memory access in high compute applications. However, ReRAM is unlikely to replace NAND Flash as it has its own manufacturing difficulties that may keep costs high. More advanced applications like real-time analytics would need something faster with high capacity, such as PCRAM.
Commercialized PCRAM products date back to the early 2000’s, but phase-change memory has finally moved out of the “emerging” category thanks to greater adoption of Intel’s 3D Xpoint non-volatile RAM (Intel Optane, see below). This technology is perhaps the best candidate for enabling massive data storage at the nanometer level while enabling extreme 3D integration. However, due to extremely precise etching and lithography required to develop something like Xpoint, the costs become prohibitive compared to other technologies. Still, IBM sees the value in PCRAM as a memory block in embedded AI systems, particularly if it can be integrated at the chip level.
Intel Optane MEMPEK1J064GAXT 64 GB Xpoint PCRAM module.
Emerging memory products are projected to produce $36 billion in combined revenues by 2030, with growth spread across embedded application areas. While it’s tempting to think of one of these technologies as being a “winner” in the memory markets, each of these technologies has its place in the embedded landscape.
BY ZACHARIAH PETERSON · THU 17 FEBRUARY 2022