Transformation touches every element of technology, flash memory being no exception. Falling prices combined with advances in flash memory have triggered a skyrocketing market for solid state disks (SSDs). They are increasingly replacing hard disk drives (HDDs) across the storage spectrum thanks to their remarkable performance, and also are finding new roles in a growing number of embedded computing applications.
Now the industry finds itself in the midst of another transformation — from 2D NAND to 3D NAND flash, which achieves higher storage densities and faster read/write operations by stacking memory cells vertically in multiple layers.
It has been a remarkable and rapid evolution.
Not that long ago, single-level cell (SLC) was considered state of the art because it was long lasting and robust. Over the past decade, multi-level cell (MLC) technology came to dominate the industrial market segment, thanks largely to falling prices and the need for higher capacities. Last year saw SSDs with 96-layer 3D (BiCS4) became commonplace. In 2020, the industry will make the leap from 96 to 128 layers, opening even greater application opportunities.
The flash area has seen similarly fast evolution to the next stage of the flash cell. As soon as 32-layer 3D flash MLC/TLC hit the market, the 64-layer 3D TLC came. And today, 96-layer 3D TLC is the standard. Early-adopter manufacturers have already received SSDs with 3D quad-logic cell (QLC). QLC is an important breakthrough because it stores four bits per cell and, thus, twice the number of states compared to TLC technology’s “mere” eight states, doubling the capacity per cell.
The NAND Flash market has never advanced as rapidly as with the advent of 3D technology. Integrating 3D NAND flash into SSDs has accelerated the pace of innovation dramatically. In 2019, SSDs with 96-layer 3D (BiCS4) became common. By mid-2020, expect them to give way to QLC-based SSDs. And 3D NAND flash should make the jump from 96 layers to 128 layers this year.
Working through decision factors for SSD selection
As with every transformation, hurdles exist. Given the multitude of SSD manufacturers and variations, it isn’t necessarily easy for users to select the right SSD for their application. Environmental conditions, size, write and erase cycles, and other factors must be carefully considered to achieve the optimal SSD performance. The following are some of the most critical factors to consider.
An abundance of different SSDs suitable for different applications are available. The first step in determining the right one to use is looking at the available capacity. However, inconsistencies in specifications make it challenging. One SSD manufacturer may specify the full flash size (i.e., how much flash is truly installed), while others may reveal only a part of the capacity and implicitly use the rest as reserve, which is also called over-provisioning (OP). The OP approach uses the additional flash capacity to, among other things, perform garbage collection to speed up internal handling and extend the SSD’s lifetime.
Here’s an example: For an SSD with flash storage area 256GB, manufacturer A might specify the full 256GB, while manufacturer B specifies 240GB, and manufacturer C specifies 200GB. With 3D flash-equipped SSDs, it is possible that the capacity actually usable by the application is smaller than the specified capacity. A common reason is that an area of flash is being utilized for internal handling.
Consequently, engineers should conduct tests in near-real field conditions to analyze the performance and lifetime of an SSD. Manufacturers often distinguish the application of the SSD using the terms “enterprise” or “client.” Enterprise SSDs are employed more for data center and server applications. They require larger flash capacity as a reserve to deliver more stable performance over a longer time period compared to client SSDs. Especially with storage arrays, engineers should make sure their designs maintain little or no latency degradation during peak loads.
Another manufacturer distinction lies in the usage data: A client SSD must be able to run eight hours a day at an average of 40 degrees Celsius, while running 24/7 at an average operating environment of 55 degrees Celsius may be required for an enterprise SSD (JEDEC standard).
Engineers should also consider capacity differences between 2D and 3D flash. 2D NAND flash (also called planar NAND) is inexpensive, but the technology has reached its physical limit with structures at around 15 nanometers (nm). This limitation can result in increased bit errors and other problems because the capacity is simply too small to hold enough electrons in the memory cells.
3D NAND flash is stacked vertically using multiple layers, resulting in higher density, better lifetime, and faster read/write operations with lower power consumption. While costlier and more complex, the cost per gigabyte is actually lower because 3D NAND packs so many vertical cells into a small size. It has far more capacity than 2D NAND within the same length and width dimensions.
The lifetime endurance or write performance endurance is an important criterion when selecting an SSD because the wrong choice can be quite costly. While endurance doesn’t play a role when using an SSD as boot medium, a very high endurance is extremely important for an SSD that has taken over the task of a data logger, for example.
Information about write performance can be found in the manufacturer’s data sheet, typically specified in terabytes written (TBW) or drive writes per day (DWPD). For example, a TBW value of 100 means that 100 terabytes of data can be written to the SSD through its lifetime. The DWPD value indicates how often the SDD can be written with the same amount of data, based on the SSD’s capacity, each day for three to five years until its lifetime is reached. So, a DWPD value of 1.9 / 5 for a 240GB SSD means it can be written every day with three times the amount of data for five years.
Several factors impact the endurance rating of an SSD, including how optimally engineers implemented the wear leveling (which distributes data writing evenly on all blocks of an SSD), write amplification factor and the W/E NAND flash cycles.
But engineers don’t rely on data sheet specs solely. Using appropriate tools, engineers need to assess the right capacity, performance and operating temperature to determine the achievable lifetime of an SSD in an application. In a use case, engineers should test the SDD in conditions similar to the real-world application. The self-monitoring, analysis and reporting technology (SMART) values from the testing tool can then be read out, or the lifetime can be displayed as a traffic-light function by means of a diagram.
For example, the lifetime can be reduced in an application where a lot of small data (<4KB) are to be written. In this case, it makes sense to pack the data first and then write it. In another example, the structure of an SSD can cause the firmware to move incoming data several times until it finally finds its place in the flash storage — a process called write amplification factor (WAF). The higher this WAF is, the more the flash cells wear out, and the lifetime declines rapidly.
Engineers can calculate the approximate write performance of the SDD once the WAF has been determined and corresponding capacity of TBW is taken from the data sheet.
Power (and power failure)
Power is a critical criterion for industrial embedded applications. Errors during data transfer are bound to happen due to power issues. To solve this, sophisticated error detection and correction (ECC) functions are implemented at every data transfer point. ECC provides potential errors with full end-to-end data path protection that secures the data transmission between host systems and NAND flash.
Engineers need to assess how stable the power supply is, while also considering the effect of a sudden failure of the supply voltage. The SSD’s internal construction and its programming with incoming data can cause certain problems, which can even lead to data loss.
New industrial SSDs now have decent protection in emergencies, thanks to integration of low-power detectors on board. If it experiences a voltage drop, the SSD controller ceases to accept any further commands and tries to save the data currently being transferred between the controller, cache and flash.
Some SSD manufacturers take the extra step of placing capacitors on the SSD board to maintain the internal voltage to safely write data that is currently in the DRAM cache. Engineers should carefully assess their application needs to ensure they design in a clean and stable power supply for the SSD.
Most SSDs today come standard in 2.5-inch packages. This size means the SSD can be used wherever an HDD was previously used, or in applications where the engineer simply wants a flexible, easily interchangeable form factor. The 2.5-inch SSD is hot swappable.
The M.2 form factor is another option, available in several versions of varying widths and lengths. The 2280 form factor, with a width of 22mm and a length of 80mm, is currently the favorite. But there are also 2230, 2242, 2260 and currently 22110 for larger capacities from 4TB. M.2 with 2280 and 2242 are produced and used in large quantities. But be aware that the M.2 is not hot-swappable like the 2.5-inch SSD.
The interface is the electrical and logical signaling between the SSD and the CPU. It defines the maximum bandwidth, minimum latency, expandability and hot-swap capability of the SSD.
Of the three basic interface options — serial ATA (SATA), serial attached SCSI (SAS), and non-volatile memory express (NVMe) — SATA is the standard in almost all systems but is reaching its limits when it comes to better performance or higher capacities. This is where SAS is comes in. If even higher speeds are required, then NVMe protocol might be the better choice. NVMe is based on the PCI express (PCIe) high-speed bus standard typically used in servers. With newer boards, this interface is used more and more in industrial and gaming applications.
SSDs typically have speeds of about 500MB/s. SSDs with NVMe, however, are two to three times faster because they are directly connected via PCIe, and bottlenecks are bypassed by SATA/SAS. Engineers should be aware that NVMe SSDs come in two versions — Gen3 x2 and Gen 3 x4 — meaning third generation with either two or four lanes to accelerate data transport speeds.
Lifetime rating is an essential factor for an SSD used in industrial embedded systems. Numerous tools on the market, such as CrystalDiskInfo, HD Tune, SSD Toolbox, SSD Fresh, HDD Health or Hard Disk Sentinel, are available (some for free) to predict lifetime of SSDs.
But if engineers need a deeper understanding about the SSD’s health, the best choice is a tool provided by the SSD manufacturers. These tools provide information about health and lifespan, as well as details on usage. Engineers can also set alerts, notifying them before the SSD reaches failure or end of life. A dashboard provides a summary of each installed SSD, displaying accurate information regarding health, temperature, capacity, wear leveling, bad blocks, etc. In other words, these manufacturer tools can almost always look deeper into the complex depths of an SSD with its interleaving of the flash and programming by means of firmware.
SSDs are experiencing a real boom and finding a place in an increasing number of innovative applications. SSDs in 2.5-inch form factor, which have now replaced smaller HDD capacities in almost all areas, play a large part in this regard. With the transition from 2D NAND to 3D NAND flash, manufacturers are producing ever-larger SSDs and flash cards, and are in extreme competition with HDD.
SSD innovation isn’t letting up. Flash memory manufacturers are already developing next-generation technologies, such as phase-change memory (PCM), resistive random access memory (ReRAM) and magnetoresistive random access memory (MRAM)/spin transfer torque (STT). It will be exciting to see which technology will assert itself in the mass market.
|Robert Herth is the senior business development manager, storage solutions, at Avnet. In this role, Robert oversees the storage and memory solutions development for Avnet Integrated. He has been with Avnet since 1993 and is based in Stutensee-Spöck (near the University city and research center Karlsruhe), Germany.|