New memory types are flooding the market. Most of them will not be successful.
There are a number of new memory types on the horizon. So why are we still using DRAM, SRAM and hard disk drives developed decades ago?
The answer is complicated. Memory, whether it’s on-chip static RAM cache or off-chip dynamic RAM—or flash storage or spinning magnetic media—is really a stack of data storage technologies that need to work seamlessly together and with other non-memory technologies. No one memory type does everything perfectly, and despite occasional pushes for a universal memory, history hasn’t shown that to be a successful strategy.
In general, the memory world is divided into non-volatile memory like flash, and volatile memory like DRAM. Volatile is slower, but it’s much more robust. With error correction it will work properly for decades. And no matter how many new memory types that have been proposed over the years, this balance hasn’t changed a whole lot.
There has been no shortage of new ideas, though. They range from biological solutions, such as amoebas and other more complex organisms and structures, to spintronics and phase-change memory, which are being studied now. And some are productized or will be soon, such as magnetic RAM, resistive RAM, ferroelectric RAM, stacked memory (Hybrid Memory Cube, 3D NAND), high-bandwidth memory, and 3D-XPoint.
All face hurdles. Some of these hurdles are obvious, others less so.
First of all, the Von Neumann architecture, on which all commercial computer architectures are based, relies on storing instruction sets and data in some version of RAM. The model was first proposed in 1945, and an entire global ecosystem has built up around that. That architecture was the master blueprint that allowed us to fit what amounts to a supercomputer in our pockets and for Moore’s Law and Dennard scaling to flourish. Whether this approach is outdated has been debated for years in computer science departments, but it’s hard to get an entire ecosystem to change direction, so most new contenders have to fit somewhere in this architecture.
Second, SRAM and DRAM will likely be around in one form or another for many years to come. Denser versions of SRAM are in the works. And most systems companies still haven’t caught up to the current capabilities of DRAM. The DDR4 spec has been out since 2014. It still isn’t widely used. There are multiple revs of this spec to improve data throughput to and from memory, as well as work underway at JEDEC for a DDR5 spec that will carry most compute architectures through the next decade. There also are different configurations of DRAM, such as high-bandwidth memory, which greatly increase the density. Coupled with advanced packaging, such as 2.5D, that also brings off-chip memory much closer to the compute elements on a chip using a multi-pipelined interconnect.
This brings up the third challenge. So while there is great excitement about new memory types and capabilities—this week’s Flash Memory Summit in Santa Clara, Calif., was packed with attendees—their targets are limited. New memory types are vying for a place between SRAM and DRAM, and between DRAM and spinning media. These are all interesting technologies, and some of them may prove commercially viable. But most of them won’t win in the long term, even if they do offer a business benefit to what’s already out there. Faster speed and lower power are particularly important in a data center, but they still need ecosystem backing to achieve economies of scale. Until that happens, it doesn’t matter how good the specs look on paper.
There is certainly an opening on the solid state drive side, because existing SSD isn’t robust enough for long-term data storage of critical information. SSD is faster than spinning media, but a magnetic disk from several decades ago can still retrieve data that is stored on it. An SSD that was introduced two years ago and heavily used since then cannot guarantee the same results.
But which solution will the market adopt? That remains to be seen. In the end, cost is the limiting factor. Any airplane manufacturer can build a supersonic jet, but no commercial airline has ever been able to keep one in service.