Every device or system that processes data needs storage and system memory for computation.
Examples of such devices or systems are remote controls for audio/video, smart thermostats, video door bells, security cameras, automotive electronics, smart phones, tablets, laptops, personal computers, air traffic systems, robotics, data center systems, …
The memories in these devices or systems can be segmented in 3 distinct categories: internet-of-things (IoT) memories, embedded memories, and high-density high-volume memories.
The memory requirements (cost, density, speed, endurance, retention, power consumption) are quite different for each of these 3 segments:
• IoT memories tend to be inexpensive, power-efficient, and low-density.
• Memories embedded in complex system chips tend to be fast, area-efficient, and medium-density.
• High-density, high-volume memories must be scalable to small geometries to be cost effective.
All functional high-yield memory technologies naturally land in one of these 3 segments. While there may be competition for market share within any segment, memories in different segments tend not to compete for market share.
High-Density High-Volume Memories
The high-density high-volume memory segment is currently dominated by DRAM (a US$50 billion market) and NAND Flash (a US$40 billion market) with these characteristics:
It is important to realize that the characteristics of DRAM and NAND Flash – as we know them today – are vastly superior to the initial commercial products introduced decades ago. Perfection is a process of gradual and continuous improvement of an initial pragmatic product that addresses a real need over many decades funded by product profits.
DRAM is super-fast, exhibits exceptional endurance, and is therefore best suited for fast system memory. DRAM, however, is expensive and volatile (the data needs to be refreshed every 60 milliseconds) and sacrifices retention to maximize speed and endurance.
In sharp contrast, NAND Flash is inexpensive with much higher bit capacity and exceptional retention, and is best suited for low-cost silicon storage. NAND Flash, however, sacrifices both speed and endurance to maximize retention.
For these reasons, the maximum silicon storage in smart phones has increased 32-fold over the past 10 years, at roughly the same cost to the consumer, while system memory has barely doubled. This illustrates that it is affordable to increase silicon storage in many products, but it is not economical to do the same with system memory.
Being limited to 2D, DRAM will likely remain expensive since silicon area largely defines cost per gigabyte. In contrast, the cost of NAND Flash is expected to decline over time thanks to 3D stacking. The cost gap between DRAM and NAND Flash will likely increase over time.
DRAM and NAND Flash fit their sweet spots near perfectly and it seems highly unlikely that a universal memory combining the best of DRAM and NAND Flash will ever exist. It is equally unlikely that any emerging memory technology will replace DRAM because its speed and endurance combination is exceptionally hard to beat. Furthermore, there is no economic justification to build a NAND Flash replacement for high-density applications while NAND Flash prices continue to erode.
However, as data processing and storage needs continue their rapid increase for mobile devices and cloud data centers, the industry needs a new non-volatile memory with attributes much closer to DRAM (because it is impossible to replace) than to NAND Flash (because it does not need to be replaced).
This vast space between DRAM and NAND Flash is therefore an opportunity for innovation. Much like the availability of inexpensive silicon storage enabled a booming new market of products that could not possibly exist without silicon storage (e.g.; smart phones), so will the emergence of new memory technologies lead to new products that we cannot even imagine today.