Despite what some experts predict, enterprise SSDs are likely to be more expensive than HDDs for some time.
This month I am going to take a look at SSD vs. HDD pricing. In my opinion, the claims by some vendors are over the top; their assertions about SSD pricing and density and HDD pricing and density simply do not match the market realities. It is time expose the real data.
I understand that SSDs do offer superior price per IOPS than HDDs (up to ~110,000 with SAS SSDs for reads but only ~40,000 for SAS SSDs for writes), but this is only part of the discussion, as people want to use SSDs for bandwidth applications.
The suggestion for combining NAND with HDDs to optimize performance is good. However, the market forecasts presented by many prognosticators that enterprise SSD prices will soon be on par with HDD prices are built on several faulty assumptions.
Also see an in-depth analysis of SSD vs. HDD Reliability.
While the idea of using flash, NAND or Solid State Drives for an intermediate step between computing nodes and disk storage is not necessarily incorrect for improving performance (Seagate is already doing this with their SSHD hybrid drives), it is built on several of flawed assumptions regarding flash storage.
First, some assume that the price of MLC NAND flash will continue to decrease at a rapid and predictable rate that will make it competitive with HDDs for bandwidth, and nearly for capacity, by 2014 or 2015. This downward trend, it is assumed, will make flash a viable alternative for large storage and to act as a memory or “buffer” to improve performance.
Second, there is a general assumption that prices for bandwidth ($/GB/s) for SSDs is much lower than for HDDs, and that enterprises will measure costs in these terms instead of capacity.
Third, there is no distinction made between flash in general, such as consumer SSDs, and enterprise storage SSDs. It is assumed that MLC NAND will not only reduce in price ($/GB) but also that it will increase in density and larger capacity drives will be developed.
Fourth, it is assumed that the quality of MLC NAND will either remain constant or increase as prices decrease and densities increase, allowing it to improve not only performance, but also reliability and power consumption of the systems it is used in.
Fifth, it is assumed that power consumption for SSDs is, or will shortly be, significantly lower than that of HDDs overall, on a per GB basis and on a per GB/s basis.
Sixth, they assume disk performance will grow at a constant rate of about 20 percent per generation and not improve.
Seventh, they assume file system data layout will not improve to allow better disk utilization.
Most of these assumptions were made in early 2012. So far they have turned out to be partially true at best and wrong at worst.
While it is true that the cost of NAND and SSDs have been going down over the past few years (and will continue to do so in general into the future), there is no proof that SSDs will match HDDs in price any time soon, especially when comparing high capacity storage options. At a $/GB measure, SSDs have dropped from upwards of $3/GB back in 2005 and 2006 down to as low as $0.67/GB for specific models in 2012, but they are still far above the cost of HDDs, which can be lower than $0.09/GB for some Western Digital and Seagate models. Even if it is assumed that SSDs continue to fall at their historical rates and HDDs fall at their historical rates, SSDs will still be about $0.15/GB in 2020 ($0.06 more than HDDs are today), while the HDDs vendors are promising may be as low as $0.03/GB.
However, several past events and the current market for SSDs means that this trend may not continue into the future or make SSDs competitive on price with HDDs for some time. First, the pricing has not always gone at a steady and predictable downward pace as current SSD prices (ranging from $1.10/GB to as low as $0.78/GB depending on vendor make and model) are actually higher than they were at the end of 2012. This is not the first time this has happened either. In 2009, SSD prices not only failed to follow the predicted 7 percent per month decreases, but actually increased as much as 7.26 percent per month for nearly four months. In fact, prices did not return to their pre-jump figures in May 2009 until a year later in May of 2010.
The current pricing trend and the hike in 2009 demonstrate that the widely discussed and frequently referenced forecast that SSDs will drop at a steady clip to catch up to HDDs in the near term is flawed.
A key component of many people’s models is the idea that disks and SSDs will not be purchased for capacity alone, but also for bandwidth. They assume that SSDs' cost for bandwidth (represented here as $/GB/s) is significantly lower than that for disk. This is not necessarily the case, especially when considering the costs for enterprise disk and enterprise SSDs.
The model is especially flawed if disk performance and utilization improves. When measured at the cost for bandwidth in $/GB/s, some consumer SSDs can get as low as $550/GB/s, but the average is between $1,000/GB/s and $1,750/GB/s.
Table1: Consumer SSD, Enterprise SSD, and HDD Comparison
*C – SSD devices are consumer SSDs, E – SSD devices are enterprise SSDs. Enterprise SSDs are colored for reference.
**All data and information was collected from manufacturer or product datasheets and publications.
Table2: Consumer SSD, Enterprise SSD, and HDD Comparison
*Prices were collected from vendor information and product prices from third party vendors and reviews. Storage and power data was collected from vendor and product datasheets and marketing papers.
**$/GB, $/GB/s, W/GB, and W/GB/s were developed based on information already available for products in this table and the previous table.
Enterprise SSDs are entirely different as the cheapest enterprise SSD had a cost of $1,303/GB/s and the average ranged from $3,200/GB/s to $4,200/GB/s. By comparison, Seagate’s Terascale 4TB HHD has a bandwidth cost of $1,438/GB/s and the Enterprise 15K 2.5” HDD has a bandwidth of about $1,000/GB/s. It is necessary to note that “bandwidth” for SSDs was measured using sequential read data, which is almost universally the best performance figure for SSDs, while HHD bandwidth was measured using average read/write times from the disk, not necessarily best performance data. As a result, real $/GB/s are probably higher for SSDs depending on the read/write workload, and will always increase with greater write workloads. As a result, SSDs do not necessarily provide an advantage or cost reduction when costs are measured in bandwidth, or $/GB/s, especially if dealing with enterprise SSDs or write-heavy workloads where wear-leveling overhead can significant increase the cost of the device to meet long term reliability requirements.
Table 3--Enterprise Hard Drive Comparison
Table 4--Enterprise Hard Drive Comparison
*Performance numbers based on average sequential read and write rates from product datasheets.
** $/GB, $/GB/s, W/GB, and W/GB/s were developed based on information already available for products in this table and the previous table.
The historical SSD cost spike in 2009 and the flat to slightly upward trend of 2013 and early 2014 demonstrate that, as with any product or service in a capitalist economic system, SSD prices are impacted as much by economic trends and developments as by technological advances. The 2009 spike was the result of increasing demand (new need for flash in smartphones, tablets, and laptops) combined with the reduction in suppliers resulting from the recession. Even the source used for this information (wikibon.org) was optimistic and assumed a continued rapid decline that turned out to be false as supplies dropped off again in 2013, with the result that prices went up and have yet to go back down. This price increase is the result of a failure to meet demand for new products, such as the Apple iPhone 5 and increasing tablet sales, and a fire at DRAM provider SK Hynix which prompted vendors to shift their flash manufacturing lines to meet a sudden shortage of DRAM.
Compared to the HDD market, which only has three main manufacturers (Seagate, Western Digital, and Toshiba), the list of SSD manufacturers is considerably longer. There are about 40 companies manufacturing and selling SSDs, ranging from computer industry giants like Intel and Seagate to much smaller start-ups. As a result, while some of the larger manufacturers (such as Intel and Samsung) can offer some support or integration benefits, the key selling point for most manufacturers remains price, especially in $/GB, since there are no significant differentiations in quality or performance. As a result, SSD prices are forced down. A future consolidation in the future could level off or even increase SSD prices considerably.
It is important to note that, while performance is a major selling point, capacity is still the dominant force. Most vendors and reviewers measure price against overall capacity, or $/GB. In fact, I found no sources in my research that ever used a price against performance figure, such as $/GB/s. Rather, they touted superior performance of SSDs as a given advantage over HDDs without any costs included. This is the result of a highly competitive market where there are few distinguishing factors between technologies aside from capacity and price point as mentioned above.
Another important market force is the SSD industry’s focus on relatively small capacity, high-performance SSDs. The main factor is the high demand for smartphones, tablets, and other small and highly mobile computing devices. In these kinds of devices, performance and small size are prioritized over capacity, and HDDs are simply too large to be practical.
This had led to an advantage for SSDs in these markets. HDDs have specific form factors, require more moving parts that are vulnerable in mobile devices, and also have preset costs for the disk, case, motor, and other necessary parts. Since they can be produced in small spaces, perform faster, and do not have to worry about load wearing, SSDs are a better option for small consumer devices. However, once you get into larger capacities and more commercial or enterprise environments, this advantage disappears as HDDs become increasingly cost efficient and SSD expenses grow rapidly.
To compete in this market, most SSD manufacturers focus on SSDs with 256 GB or less, while a few manufacture SSDs with 512 GB, and even fewer manufacture anything above 1 TB. As a result, while prices for SSDs in the 256 GB and below range continue to decrease, SSDs of 512 GB and above have leveled off, and SSDs with 1 TB or more have prices described as “astronomical.”
This price difference means that, while small and mobile devices continue to increase SSD usage, large storage and enterprise level systems will continue to use HDD or tape storage for the foreseeable future.
There are few manufacturers actually making SSDs at 1 TB capacities and above, and there are even fewer actually manufacturing what would be considered enterprise SSDs. The majority manufacture consumer SSDs.
Consumer and enterprise SSDs have several key differences. While consumer SSDs focus on lower cost, enterprise SSDs focus on performance and quality. As a result, consumer SSDs tend to be built with a SATA interface common in laptops and PCs, while enterprise SSDs are built with either SAS (for reliability) or PCIe (for performance) interfaces. In addition, low-quality MLC and TLC are used in consumer SSDs that strictly adhere to a 2.5” form factor to fit in consumer demands. Enterprise SSDs use either SLC (significant downtrend over the last year in production) or more commonly Enterprise MLC (eMLC) with 2.5”, 3.5” and special large form factors to meet demands.
The resulting difference in costs is considerable. Consumer SSDs tend to range from $300 to $650 depending on make, model and capacity, while enterprise SSDs can range from $2,250 to $16,000 depending on make, model and capacity. This range means that, while consumer SSD prices are relatively close, predictable and decreasing (especially on a $/GB basis), enterprise SSDs designed for database or supercomputer applications vary widely both in overall costs and $/GB.
The differences and trends in performance are even more significant than cost. Consumer SSDs all seem to have maximum sequential read performances of about 500 to 550 MB/s and write performances up to 540 MB/s while random 4k IOPS never really exceed 100,000 for reads and range from 25,000 to 90,000 for writes.
On the other hand, enterprise SSDs using SAS have sequential read performances ranging from 430 MB/s to 925 MB/s and writes from 175 MB/s to 595 MB/s with random 4k IOPS (I/O request per second) as stated above. Enterprise SSDs that use a PCIe interface have by far the best performance with individual SSDs that usually have sequential read performance in excess of 3 GB/s and write performance between 630 MB/s and, in one special instance, 3.2 GB/s. Random 4k IOPS for PCIe enterprise SSDs could reach up to 750,000 for reads and 275,000 for writes, with most having IOPS of between 100,000 to 400,000 for reads and 65,460 to 200,000 for writes.
The consumer SSD products have very little differentiation among them and actually share similar performance numbers as well as price, while enterprise SSDs vary greatly. It should be noted that no real application can really take advantage of SSDs that support over 100,000 IOPS.
The status of consumer SSDs for both price and performance suggest a saturated market with little technological differentiation where total prices and price for capacity ($/GB) dictate the market, artificially lowering prices. The wide disparities in enterprise SSDs show a different market where products are specialized for specific workloads, environments or demands.
The general drop-off of write performance from consumer to enterprise SSDs against read performance strongly suggests that enterprise SSDs are incorporating additional error correction code (ECC) and other measures to ensure that data is written to the drive more reliably.
In fact, only one drive, the Super Talent RAIDDrive II Plus, actually offered greater write performance than read performance as it was designed for write heavy workloads and given a special caching system. This is also the only instance where write performance saw a significant increase from a predecessor system (the original RAIDDrive), otherwise, consumer SSD write performance has seen only slight increases, while enterprise SSDs have either remained constant or even decreased by as much as 50 percent to maintain data reliability.
These differences are important to consider when making forecasts regarding the use of SSDs in enterprise environments. The enterprise SSD environment is less predictable—and much more expensive—than the consumer SSD market generally used for SSD forecasts.
Even if economic forces are favorable to continuing price reductions of SSDs and NAND flash, a 2012 study by Microsoft Research (PDF) has found that a dilemma arises when trying to increase density and reduce cost of SSDs. The study looked at 45 flash chips from six different manufacturers and found that, as density increases, bit error rate (BER) and performance decrease. This is because the number of ranges of electrical charges necessary to store data on a single cell increase as densities increase.
The researchers found that, as feature size decreases (increasing density), bit error rates increase. While the SLC and MLC chips with cells that had feature sizes of between 80 and 60 nanometers (nm) usually had BERs of 1e-08, those with feature sizes of 40 nm had BERs at or below 1e-07, and the TLC chips with feature sizes of 20 nm had BERs of, at best, 1e-03.
In addition, researchers also found that increasing density also increases read and write latencies. NAND chips with feature sizes above 64nm had read latencies of 20us or less and write latencies of 0.5ms or less, while those with feature sizes of 32nm or less had read latencies between 20us and 60us and write latencies between 0.5ms and 2.5ms.
This leaves SSD and NAND manufacturers with a choice among density, cost, reliability, and performance. In any scenario, at least one of these four must be sacrificed to improve the others. This means that, even if SSDs can achieve cost parity with HDDs, it will be at the expense of reliability, density or performance. In fact, as discussed above, enterprise-grade SSDs already sacrifice write performance, cost and even density to address the threat of reduced reliability and data integrity and have built non-2.5” form factor configurations and added special coding or technologies to meet reliability and performance demands, resulting in more costly products.
The last key assumption in many arguments is that SSDs present a much more efficient alternative to disks in regards to power consumption and that they will become more important because database and supercomputing systems will hit a power consumption barrier of about 20 MW. Some point to the fact that SSDs at idle save far more power than HHDs constantly spinning, meaning that SSDs offer a cheaper option in the long run as they will use up less power.
The reality is that, currently and into the near future, this issue is not as clear as some would like to think.
The Seagate Enterprise 15K 2.5” form factor HDD and Terascale HDD have power consumption needs of 1 W and 6.5 W per drive, respectively. However, SSDs are far more varied. Consumer SSDs, designed for laptops or tablets, often have power consumptions of between 0.1 and 1.5 W per drive, however enterprise SSDs can range from 3 W to 30 W depending on make and model with most falling between 3 W and 10 W.
It is important to note that the power consumption measures used were the HDD and SSD idle rates, the benchmark that should make SSDs more efficient, not maximum power consumption during IO bursts. In fact, if you compare a measure of W/GB of storage, HDDs are still more efficient per GB of storage than all but a few consumer SSDs designed for laptops.
In regards to bandwidth, measured in W/GB/s, consumer SSDs are significantly more efficient than enterprise options, however the Seagate Enterprise 15K HDD is actually competitive with most of the enterprise SSDs. Only two of the high performance SSDs have efficiencies that are more than 1 W/GB/s less than the Seagate Enterprise 15K HDD efficiency of 5 W/GB/s.
Given this information, the power efficiency gains of using SSDs instead of HDDs that some suggest are minimal if not non-existent in enterprise environments.
While the idea of integrating NAND flash technology into a storage solution to improve performance is not necessarily a bad idea, the ability to do so on a large scale is built on several faulty assumptions regarding SSDs.
First, the optimistic price trends presented ignore real price trends that have not always been reliably downward and have, at times, gone upward.
Second, the assumption that SSDs are cheaper in terms of bandwidth ($/GB/s) and that users will measure costs in this way may be true of the consumer market, but enterprise SSDs are actually more expensive on a $/GB/s basis. No vendor or third party has put forward costs in terms of bandwidth, such as the $/GB/s measure used in this paper. In addition, the assumption that prices will always decrease ignores key market forces, such as supply availability, competition and lack of demand or need for large storage options that have made SSD pricing far more unpredictable than presupposed.
Third, some make no distinction made between consumer SSDs and enterprise SSDs despite the fact that consumer SSDs and enterprise SSDs exist in entirely different market spaces, resulting in significant differences in interfaces, performances, costs and trends between them.
Fourth, these SSD assumptions ignore research that suggests that either reliability, performance, density or a combination of the three must be sacrificed to continue to reduce prices while developing NAND technology in the near future.
Fifth, the assumption that SSDs will require less power today and into the future because they have no moving parts and can idle ignores the complexity of SSD market needs and increased demands for power and applications on enterprise SSDs to ensure reliability and performance.
In the end, flash technology and SSDs cannot yet replace HDDs as primary storage for enterprise and HPC applications due to continued high prices for capacity, bandwidth, and power as well as issues with reliability that can only be addressed by increasing overall costs.
Last but certainly not least is a graphic from EMC:
At least for the foreseeable future, the cost of flash compared to hard drive storage is not going to change.
Photo courtesy of Shutterstock.