New non-volatile memory technologies poised to reshape system and application design

Business and IT execs can be excused for nodding off whenever the topic turns to component technologies and system internals, but occasionally the ostensibly big new thing turns out to be broadly significant.

As I recently discussed, Moore’s Law has hit its golden years and put CPU technology on a decelerating performance curve, so it’s easy to assume that the best days for the silicon part of Silicon Valley are behind it.

Yet progress in material science and semiconductor physics marches on as a new generation of engineers and entrepreneurs solve what once seemed intractable problems, making old ideas new again. Such is the case with resistive storage technologies such as ReRAM, MRAM and phase-change memory that have been the topic of sporadic research and breathless (and unfulfilled) promises for three decades. Yet we've finally reached the point where critical roadblocks of scalability and manufacturability are being overcome.

As a device physicist and DRAM designer early in my career, I’ve seen all the predictions about the putative superiority of alternative memory technologies. Over the years my healthy skepticism turned into downright cynicism as one claimed revolutionary announcement after another came to naught. But even a curmudgeon can be convinced by evidence and several developments have convinced me that the era of viable resistive alternatives to DRAM and flash is here.

Unlike conventional memory, where the basic unit of storage is charge on a capacitor (in the simplest form, it's either there or not, 0 or 1), resistive memory works by altering the resistive properties of a material to produce one (or more) states which again can represent a 0 or 1. Several companies are poised to deliver data center and military-grade storage devices with capacities large enough to be practically useful.

The significance isn’t just in being operationally different, but in performance characteristics that place resistive memory squarely between DRAM and flash, thus presenting some enticing possibilities for system and application design.

Complementary contenders: Intel-Micron have company

The best known of the new memory technologies is 3D XPoint, a joint effort of Intel and Micron. Although the companies haven’t offered details beyond admitting that it’s “not based on electrons”, 3D XPoint is thought to use some sort of phase change material that can take on several different resistive states in response to electrical signals.

Regardless of the material science, its benefits as a storage medium are significant: providing non-volatile storage like flash with 1,000 times the durability, hundreds of times the overall performance and comparable chip-level density (i.e. 8-10x more compact than DRAM).

Intel sees 3D XPoint occupying a new tier in the memory hierarchy as a non-volatile adjunct to DRAM for CPU memory. Given 3D XPoint's density advantage, combining the two could enable much larger in-memory systems for databases and business analytics while obviating the need for a separate flash or disk tier for hot-data persistence.

Intel hasn't provided devices for independent benchmarking, however its unverified results are impressive. Still, it's unclear how well 3D XPoint will work as a drop-in replacement for DRAM versus as an accelerator for flash.

That second scenario is where another storage alternative holds promise. Everspin, a pioneer in magneto-resistive memory (MRAM) that has been working with the technology for nearly a decade, recently began shipping 256Mbit non-volatile devices that are comparable in performance to DRAM, i.e. 5 orders of magnitude faster than flash, yet use the same DDR3 and 4 memory interface and provide the same ability to handle virtually unlimited data writes.

Admittedly, this is one-sixteenth the size of current generation DRAM and smaller still than 3D NAND chips. However Everspin CEO Phil LoPresti says 1Gbit parts are coming soon and that the technology, which uses a standard CMOS process for logic and control circuitry, can scale much further.

MRAM's benefits to system designers as a flash alternative are significant, chiefly due its 20-year data retention, unlimited endurance with no wear out mechanism (i.e. unlimited writes), symmetrical read and write cycle times and reliability across a wide temperature range. These attributes make MRAM look more like DRAM, whereas 3D XPoint still exhibits write latency and endurance limitations like flash.

Turbocharging application performance

The implications for application performance are particularly intriguing when considering how these four memory technologies might be combined. DRAM will always be the fastest and most general purpose system memory while flash has the density and resulting price/performance advantage to displace hard disk for more and more workloads.

Yet, inserting a layer of MRAM and 3D XPoint in the storage hierarchy would provide applications a huge boost in both read and write performance: MRAM for the write caching a flash subsystem and 3D XPoint for a much larger read cache.

Imagine a storage system capable of 10M IOPs, which Everspin claims is easily achievable, yet with the persistence of disk. The performance improvement for journaling file systems, structured and unstructured databases, associated analytics software and various data streaming applications would be profound.

Being plug-compatible with DRAM, a mix of DRAM, MRAM and 3D XPoint could yield servers with many times the system memory at much lower cost than today's in-memory DRAM designs.

My take

We're still early in the commercialization and maturation of alternative resistive memories so its ramifications on application performance and software architectures are still more theoretical than actual.

However software developers should begin to consider how they might design applications in a world where they have access to much larger pools of lightning fast, byte-addressable persistent memory.

Likewise, IT leaders should start thinking about the types of applications and services they could provide if databases and other storage systems delivered several orders of magnitude faster performance than even today's best all-flash arrays.

The Machine from HP is just such an experiment, nor by no means the only one, to see what's possible by essentially eliminating the logical and performance boundaries between system memory and storage, an eventuality HP believes will fundamentally change computer architecture.

No, DRAM and flash aren't dead, but they are about to get some powerful turbochargers.