There’s been quite a lot of innovation in IT storage media over the years. My own journey, like many of us in the field, is a long one. It started with something like magnetic tape, moved to hard disk drives (HDDs) and then flash-based solid-state drives (SSDs). The technical attributes of any storage media are ultimately what help to drive IT adoption. There are several factors which make any storage media a market success, but the underlying technology, and how it is accessed, is essential. Let's take a look at a few technical characteristics of storage media that matter to IT.
The amount of space that is available needs to match or exceed the data footprint. This can be pure physical or raw storage capacity, or simply what the device presents to the OS or storage system. Compression and other data reduction techniques allow “logical” capacity to exceed physical capacity. Data protection can have the opposite effect, reducing logical capacity due to the need for redundancy to protect customer data from media failure.
Many times the impact of storage performance is overlooked when considering the overall speed of the deployment. For most use cases speed is typically measured as time to first data, also known as access time, latency, or response time. This can be sustained drive throughput or the amount of time between active accesses. Similar metrics exist for writing data around when the data is persistent, the most valuable part of storage.
All storage requires power to do active work. It's typically measured at idle and during read/write I/O workloads. As compute used for AI (CPU and GPU) consume more of a data center’s energy budget, IT has had to get creative with power used by other parts of the infrastructure.
When data is written to storage it must be able to be read back. There are several measurements that are accepted today that include bit error rate (BER) and mean time between failure (MTBF). Some additional HDD and SSD specific features now include drive writes per day (DWPD) and total bytes written (TBW) to represent storage longevity.
There have been many transitions in the primary data center storage solutions over the years. What started with punch cards moved to tape, then to HDD, and now SSD. It’s the transition from HDDs to SSDs that’s remaking the enterprise storage market today.
The era of data center storage started primarily with tape-based solutions. Tape died away in favor of HDDs with the advent of the first room-sized solutions. HDD went on to dominate the enterprise storage media market in partnership with tape for decades afterwards.
Interestingly, tape was all but pronounced dead in the 70s and 80s. However, tape is still being used for a significant amount of archived data, known as “cold” or “frozen” storage. It exists for use in cloud and/or HPC shops that still have need of the exabytes (EBs) to yottabytes (YBs) of data that persist in the world today.
To understand this, a discussion on the unique abilities of these different storage solutions seems in order. First off, there is a need to view the basics of tape, HDD, and SSD in a somewhat unique way.
Some characteristics of tape that caused its decline include:
Aside from archive use, another reason tape is still used in cloud/HPC data centers is because of its energy consumption and longevity. Essentially, tape uses 0W of power when unmounted and media longevity is >10 yrs.
HDDs are also being relegated to more niche markets. Like tape, similar problems are impacting disk. These are heavily related to moving parts, like spindles, platters, and motors. Even interfaces for HDD are a challenge:
Currently HDDs are still shipping large volumes and will continue for some time, but like tape, these are primarily going to hyperscalers and other customers with EBs to YBs of data that requires archiving, rather than active storage.
Shifting gears, SSDs and the flash technology market are booming, across the board. Why?
New AI low latency access methods and interface protocols target SSDs: As the underlying media in SSDs is semiconductor technology, access can be at electronic speeds and can supply response times in microseconds. Sustained throughput for SSDs is generally 3 GB/sec to 4 GB/sec or more. SSD throughput is a function of the parallelism in a flash IC, the number of flash ICs in a package, and the number of flash packages in an SSD. This allows drive throughput to be increased much easier than for HDD or tape. Learn more about packaging in our article: How Solidigm Drives the Foundation of Innovation in SSD Packaging Technology
Increasing SSD capacities have challenges but are doable: There’s essentially three ways to increase SSD physical capacity, media scaling, package scaling, and bits-per-cell scaling. The first SSDs used single (1) level cell (SLC). Today we see triple (3) TLC and quad (4) QLC devices. Current flash vendors are shipping devices that have been stacked (3D NAND) up to and beyond 200 flash layers per die. Given all its scaling, nailing down Gbpi2 for flash is a moving target but typically SSDs are 15 Gb/mm2 (~9,600 Gbpi2) or higher. And as mentioned earlier, flash can come in more than one IC per package. Solidigm™ D5-P5336 QLC SSD supports 122.88TB of physical storage capacity.
SSD ecosystem is expanding: NAND flash is based on semiconductors and allows for more companies to work on advancing semiconductor technology and drives at least a part of SSD’s ecosystem. While there are several key vendors, 5 and growing, the ecosystem of support organizations is larger as well.
With the fastest access times (microseconds) of any storage media, the highest capacities both total and Gbpi2, the most throughput at >4GB/sec per drive, the lowest energy costs per storage (TB/W), and the fastest expanding technology ecosystem, it is no wonder SSDs are taking over the IT storage market today.
It’s important to realize that applications are one of the primary drivers of IT storage media adoption. One reason for tape’s fall from data center grace is that applications evolved to start using random access memory (RAM). Before that system, users were required to batch-in/batch out, which drove sequential access. The need for random access created a window of opportunity for HDDs to become the main type of storage. As such, changes in enterprise applications and their access patterns have an outsized impact on IT storage media adoption.
The biggest change to hit enterprise IT applications today is AI. Its emergence over the last decade took many enterprises by surprise but by now, AI adoption is full steam ahead.
But it’s not just AI deployment that’s driving enterprise application change.5
Some of the key factors include:
Having low power, faster, and higher capacity SSD storage that can keep up with these trends has become even more vital.
Storage media transitions happen infrequently in the data center. Sometimes it’s difficult to predict, but the trends can be identified if one looks hard enough.
SSDs have been around for decades. However, unlike previous efforts to make applications stay focused on access patterns, the storage media in SSDs allow AI enterprise access to be tailored for methods/interface protocols to coincide with SSDs, leaving HDDs far behind. Moreover, SSDs’ raw capacity is running away from all other storage media and given SSDs’ low energy use, they are rapidly becoming the storage media of choice for any leading-edge enterprise adopting AI.
Tape still exists ~50 years even after it was pronounced dead. HDDs still exist but we, as an industry, are almost ready to put it on archive life support. SSDs are really the only storage media solution for keeping up with the accelerating pace of application change at the enterprise level, especially as companies shift to incorporate more AI into their data center workloads.
Scott Shadley is Director of Leadership Narrative and Evangelist at Solidigm, where his focus is on efforts to drive adoption of new storage technologies, including computational storage, storage-based AI, and post-quantum cryptography. Scott brings over 25 years of experience in the semiconductor and storage space, where he has played a key role in both engineering and customer-focused roles.
