TECHNOLOGY IN CONTEXT
The Expanding Roles of Nonvolatile Memory
The Expanding Role of Non-Volatile Memory in High-Performance Embedded Architecture
The answer to the big question, “What will be the next non-volatile memory to replace NAND flash?” is currently unclear, but due to the limitations of NAND flash (endurance, reliability, speed) when compared to DRAM, it is likely a new non-volatile memory technology will evolve.
ADRIAN PROCTOR, VIKING TECHNOLOGY
Page 1 of 1
Over the past 20+ years, there have been numerous memory technologies brought to market with varying degrees of commercial success. Among these are static RAM (SRAM), pseudo static RAM, NOR flash, Eprom, EEprom, DRAM and NAND flash. Generally speaking, these “memory” technologies can be split into two categories, volatile and non-volatile. Volatile memory will not retain data when power is turned off; conversely, non-volatile memory will. The two dominating memory technologies in the industry today are DRAM (volatile) and NAND flash (non-volatile). Figure 1 summarizes memory technologies as emerging, niche and those in mass production.
Memories categorized as mass production, niche application or emerging technology.
The memory technologies that dominate the computing industry today are, DRAM and NAND flash, but it should be noted that both of these technologies have their pros and cons (Table 1). DRAM delivers the highest performance (latency / speed), with practically infinite endurance, yet it is volatile and has much lower capacity points than other memories such as NAND flash. On the other hand, NAND flash scales to high capacity, is non-volatile and relatively cheap ($//Gbit), but it is significantly slower than DRAM. Additionally, endurance and data retention are getting worse as process geometries continue to shrink, meaning that for write-intensive enterprise applications, NAND flash, in the long term, may not be an optimal memory technology.
There is much discussion in the industry as to what new universal memory technology or technologies will materialize as real contenders to displace either or both NAND flash & DRAM. Some of these newer emerging technologies include: Magnetic RAM (MRAM), Ferroelectric RAM (FRAM), Phase Change Memory (PCM), Spin-Transfer Torque RAM (STT-RAM) and Resistive RAM (ReRAM) or memristor. See sidebar “Emerging Memory Technologies Overview,” p. 22.
FRAM, MRAM and PCM are currently in commercial production, but still, relative to DRAM and NAND flash, remain limited to niche applications. There is a view that MRAM, STT-RAM and ReRAM are the most promising emerging technologies, but they are still many years away from competing for industry adoption. Any new technology must be able to deliver most, if not all of the following attributes in order to drive industry adoption on a mass scale: scalability of the technology, speed of the device, power consumption better than existing memories, endurance, densities better than existing technologies, and finally cost—if the emerging technology can only manage one or two of these attributes, then, at best, it is likely to be resigned to niche applications.
So the answer to the question, “What will be the next non-volatile memory to replace NAND flash?” is almost certainly that a new non-volatile memory technology will evolve. However, it probably will not replace the current mainstream memories for at least the next 5 to 7 years. However, non-volatile DIMMs, such as the ArxCis-NV from Viking Technology, can enable increased application performance and far improved power failure / system recovery, when compared to current implementations (Figure 2).
The Viking Arx-Cis-NV fits into a socket in main memory but has power backup via a supercapacitor, making it a non-volatile DRAM.
What Is a Non-Volatile DIMM?
A non-volatile DIMM is a module that can be integrated into the main memory of high-performance embedded compute platforms, such as AdvancedTCA Blades; perform workloads at DRAM speeds; yet be persistent and provide data retention in the event of a power failure or system crash.
A non-volatile DIMM is a memory subsystem that combines the speed and endurance of DRAM together with the non-volatile data retention properties of NAND flash. This marriage of DRAM and NAND technology delivers a high-speed and low-latency “non-volatile / persistent” memory module. Designed from the ground up to support unlimited read/write activity, it performs at fast DDR3 speeds and can sustain itself in the face of host power failure or a system crash.
What makes these modules different from standard DRAM modules is that in the event of a power failure or system crash, the data in the NV-DIMM is securely preserved and available almost immediately upon power being restored to the host system, much as in the case of suspend/resume. If performance is critical to business success and if minimizing downtime is an important issue, then NV-DIMMs will be extremely valuable wherever the application bottleneck is storage and I/O, and where downtime costs money.
With the recent surge in use of SSDs in high-performance compute environments, and those architectures also utilizing caching software (auto-tiering), many applications have enjoyed significant performance improvements. Therefore the SSD and software bundle has significantly improved the memory/storage gap, which has helped alleviate I/O bottlenecks (Figure 3).However, the fact remains that NAND flash SSDs are best suited for read applications, not write-intensive ones. Thus, these intelligent caching software solutions, when paired with SSDs, will utilize other system resources to ensure performance & reliability, i.e., CPU, DRAM and HDDs.
Indeed most SSD caching software will utilize the host system’s standard DRAM for intensive write activity to preserve the SSD and also keep the write cache in DRAM. That means this critical data is susceptible to loss in the event of a system crash or power failure. A common work-around to protect the data is to “checkpoint” the write buffer out into slower block-based storage, namely disk, at regular intervals. But “checkpointing” naturally has a negative impact on the I/O performance.
With NV-DIMMs integrated into the host system, data-centric and write-intensive applications will enjoy a significant increase in performance. In addition to the benefits of increased application performance, recovery is another area of significant benefit.
Should the system experience a power outage, one of two scenarios will occur. Either the power failure will cause a catastrophic loss of the “in-memory state,” or the backup power supplies will enable the appliances to transfer this data held in main memory out to disk. The entire state—which could be hundreds of gigabytes of DRAM— must be saved to a storage back end. Both “saving” and “recovering/reconstructing” this amount of data, across multiple servers, will be extremely slow and place a very heavy load on the storage infrastructure, resulting in severe I/O bottlenecks.
When utilizing non-volatile DIMMs, in the event of power loss or if the system crashes unexpectedly, the NV-DIMMs allow the system to recover their “in-memory state” almost instantaneously without putting any load on the storage back end—in a sense, making the failure appear as a suspend/resume event. This means that business-critical applications such as online transaction processing (OLTP) can be up and running again in a matter of minutes rather than hours and without the need for uninterruptable power supply (UPS) intervention.
The Role of the Supercapacitor
In order for a non-volatile DIMM to perform its task, a small energy source is required to ensure 100% data security on system failure. Supercapacitors are an appropriate technology for use in this environment, primarily because they provide a superior solution when compared to batteries. Any technology, when new and relatively unknown, will encounter questions about long-term reliability and capabilities. There are a number of reasons why supercapacitors are suitable for use in this type of application.
Supercapacitors are highly efficient components with high current capability. Their efficiency—defined as the total charge removed divided by the total charge added to replenish the charge removed—is greater than 99%, even at very high currents, meaning that little charge is lost when charging and discharging the supercapacitor. Since supercapacitors are designed with a very low equivalent series resistance (ESR), they are able to deliver and absorb very high current. The inherent characteristics of the supercapacitor allow it to be charged and discharged at the same rates, something no battery can tolerate. In battery-based systems, you can only charge as fast as the battery will accept the charge.
Since supercapacitors operate without relying on chemical reactions, they can operate over a wide range of temperatures. On the high side, they can operate up to 65°C, and withstand storage up to 85°C, without risk of thermal runaway. On the low side, they can deliver power as cold as -40°C.
Determining battery state of charge (SOC) and state of health (SOH) is a significant consideration for robust battery systems, requiring sophisticated data acquisition, complex algorithms and long-term data integration. In comparison, it is very simple to determine the SOC and SOH of supercapacitors. At the same time, the energy storage mechanism of a supercapacitor is capable of hundreds of thousands of complete cycles with minimal change in performance. They can be cycled infrequently, where they may only be discharged a few times a year, or they may be cycled very frequently.
Life cycle and maintenance are also important factors. The energy storage mechanism of a supercapacitor is a very stable process. It is capable of many years of continuous duty with minimal change in performance. In most cases, supercapacitors are installed for the life of the system. In addition, supercapacitors cannot be over charged/discharged, and can be held at any voltage at or below their rating. If kept within their wide operating ranges of voltage and temperature, there is no recommended maintenance.
The architecture of NV-DIMMs provides a full interconnect on-module that will independently allow transfer of data between the DRAM and the flash without contention for other I/O or CPU resources. Ultimately, applications can rely on the high-speed memory (DRAM) to be “persistent” and need not slow down to “checkpoint” or consume other system resources. The NV-DIMM delivers value that far surpasses a simple DRAM, DIMM and SSD architecture; it is greater than the sum of the two technologies used.
Emerging Memory Technologies Overview
MRAM: Magnetic RAM
MRAM is a non-volatile memory. Unlike DRAM, the data is not stored in electric charge flows, but by magnetic storage elements. The storage elements are formed by two ferromagnetic plates, each of which can hold a magnetic field, separated by a thin insulating layer. One of the two plates is a permanent magnet set to a particular polarity; the other’s field can be changed to match that of an external field to store memory.
Vendors: Everspin (Freescale spin-off), Crocus Technology
STT-RAM: Spin-Transfer Torque RAM
STT-RAM is an MRAM, which is non-volatile, but with better scalability over traditional magnetic RAM. STT is an effect in which the orientation of a magnetic layer in a magnetic tunnel junction or spin valve can be modified using a spin-polarized current. Spin-transfer torque technology has the potential to make possible MRAM devices combining low current requirements and reduced cost. However, the amount of current needed to reorient the magnetization is at present too high for most commercial applications.
Vendors: Samsung, SK-Hynix, Renesas, Toshiba, Everspin, Crocus Technology
PCM: Phase Change Memory
PCM is a non-volatile random access memory. It utilizes the unique behavior of chalcogenide—a material that has been used to manufacture CDs—whereby the heat produced by the passage of an electric current switches this material between two states. The different states have different electrical resistance, which can be used to store data. It is expected PCM will have better scalability than other emerging technologies.
Vendors: Micron, Samsung
ReRAM: Resistive RAM
ReRAM is a non-volatile memory that is similar to PCM. The technology concept is that a dialectric, which is normally insulating, can be made to conduct through a filament or conduction path formed after application of a sufficiently high voltage. Arguably, this is a memristor technology and should be considered as potentially a strong candidate to challenge NAND flash.
Vendors: SK-Hynix, HP, NEC, Panasonic, Samsung
Foothill Ranch, CA.