Sign in with
Sign up | Sign in
Intel SSD 710 Tested: MLC NAND Flash Hits The Enterprise
By ,
1. Intel On Enterprise Storage: No More SLC; Meet HET MLC

You'll find some of the highest-end computer hardware in the largest data centers and supercomputing labs; stuff that would blow your mind. When it comes to the high-performance storage in those environments, SSDs based on single-level cell (SLC) memory elements are often favored for their great performance, power, and reliability characteristics.

In the early days of solid-state storage, multi-level cell (MLC) NAND SLC-based SSDs were deemed unsuitable for the write-intensive nature of many server workloads. The technology weathers fewer program and erase (P/E) cycles before deteriorating. Moreover, MLC achieves slower write speeds than SLC-based cells. And, in the process, MLC memory uses more power (an important consideration in a data center potentially playing host to thousands of drives).

Big Deployment: 5000 X25-Es at SoftlayerBig Deployment: 5000 X25-Es at Softlayer

As a result, many IT managers look to SLC-based drives for applications where data security and speed are of the utmost importance. Specifically, Intel's X25-E (first reviewed back in 2009: Intel’s X25-E SSD Walks All Over The Competition) is the benchmark by which other enterprise-class SSDs are measured. 

There are a couple of problems, though. First, as its name suggests, an SLC memory cell only stores one bit of data. Compute-quality MLC stores two. Right off the bat, you can see that multi-level cell technology is what facilitates the higher capacities many SSDs enjoy today. Intel's X25-E, in comparison, topped out at 64 GB. The other issue is price. That same 64 GB flagship sells for as much as $800, more than $12 per gigabyte of storage. 

Clearly, if a manufacturer could figure out a way to push the benefits of MLC-based NAND into the enterprise without compromising data integrity, there would at least be a compelling reason to start slinging larger SSDs together in RAID, or using them singularly as caching devices in a tiered storage subsystem, right?

Well, Intel certainly thinks so. The company is discontinuing the X25-E altogether in favor of a new SSD 710, representing a shift from expensive SLC to more accessible MLC memory.

Despite the fact that Intel's new data center drive takes the MLC route, the company says it delivers a different experience than the mainstream SSD 320. The NAND found in Intel's new enterprise SSD is dubbed "High Endurance Technology (HET) MLC", which tries to balance the capacity benefits of MLC and write endurance of SLC memory.

The move away from SLC naturally involves some compromise. However, from a big-picture approach, it makes sense. When you combine the technical barriers to SLC production and factor in economies of scale, at the same density, SLC NAND commands a price premium 4x higher than MLC, according to data from iSuppli. An MLC-based drive is going to be much more accessible to cost-conscious SMBs and larger data centers alike.

Cost
Market Price (Debut)
Price Per GB
Intel X25-E 32 GB
$460
$14.38
Intel X25-E 64 GB
$900
$14.06
Intel 710 SSD 100 GB
$679
$6.79
Intel 710 SSD 200 GB
$1299
$6.50
Intel 710 SSD 300 GB
$1999
$6.63


Oh yeah. Look at the difference in price per gigabyte. The X25-E debuted at about $14/GB. More than two years later, just before Intel announced it was discontinuing the X25-E, it had only dropped to about $11/GB. That's still a 40% price premium over this new SSD 710, though. But there's more to the story than just dollars and cents.

Consumer-oriented SSDs are still hovering near $2/GB. So, an MLC-based SSD priced at $6.50/GB should still rightly raise some eyebrows. However, the HET MLC found in Intel's SSD 710 is purported to offer write endurance 30 times greater than the cheaper consumer-grade MLC. So, if we assume that 25 nm MLC NAND is rated at 3000 P/E cycles, HET MLC should come close to 90 000 cycles. That SLC-like write endurance rating is intended to assuage the fear of IT managers now unable to purchase the X25-E and faced with SSD 710. A substantially-lower price per gigabyte, meanwhile, is designed to attract the contingent of folks stuck using magnetic storage because they couldn't stomach the premium on SLC memory for their mission-critical data.

2. Inside The SSD 710: Something Old And Something New

Intel's PC29AS21BA0 NAND ControllerIntel's PC29AS21BA0 NAND Controller

Although this is technically a brand new SSD family, everything under the SSD 710's hood is pretty darned familiar. In fact, this enterprise-oriented drive is a mirror image of the consumer-oriented SSD 320. Both employ a ten-channel architecture based on Intel's PC29AS21BA0 controller. The difference, of course, is that the 710 SSD uses HET MLC, which is responsible for a scaled-back write speed rating. Beyond the controller and NAND, the SSD 320 and 710 share the ability to apply AES-128 encryption and protect data during power outages through the use of on-board capacitors.


Intel SSD 320 (G3)Intel SSD 710
Capacities120/160/300/600 GB
100/200/300 GB
NANDIMFT 25 nm MLC, ONFI 2.2
IMFT 25 nm HET MLC, ONFI 2.2
Cache
64 MB DRAM, 166 MHz
64 MB DRAM, 166 MHz
Sequential Read270 MB/s270 MB/s
Sequential Write220 MB/s210 MB/s
4 KB Random Read39 500 IOPS38 500 IOPS
4 KB Random Write23 000 IOPS2700 IOPS
SecurityATA Password + AES-128ATA Password + AES-128


Put a 300 GB SSD 320 and a 200 GB SSD 710 next to each other; it's difficult to distinguish one from the other.


Intel SSD 710
200 GB
Intel SSD 320
300 GB
Market Price (Debut)
$1299
$549
Price Per GB
$6.50$1.83
Raw Flash
320 GB320 GB
IDEMA Capacity
200 GB300 GB
User Accessible
186.31 GiB279.46 GiB
Over-Provisioning
40%6.66%


Like the 300 GB SSD 320 in our lab, Intel's 200 GB SSD 710 has 20 NAND packages, each one adding 16 GB to the drive's capacity. But while each SSD's PCB is aesthetically identical, the company's enterprise offering employs 40% over-provisioning. That's the highest we've seen from any product. This is a critical component in the adoption of MLC in the enterprise space, though, as increasing over-provisioning decreases write amplification, which in turn positively impacts the drive's endurance.

While 40% sounds like a lot, Intel recommends even more over-provisioning for write-heavy applications. If it were to set aside an additional 20% of the drive's NAND flash, write endurance would increase by another 50% or so. Combining HET MLC and exorbitant amounts of over-provisioning allows the SSD 710 to achieve an endurance rating 33x higher than the consumer-flavored SSD 320.

Write Endurance
(20% Extra Over-Provisioning In Parentheses)
NAND
Capacity Point
Endurance Rating
Intel SSD 320
MLC
160 GB
300 GB
600 GB
15 TB
30 TB
60 TB
Intel SSD 710
HET MLC
100 GB
200 GB
300 GB
500 TB (900 TB)
1.0 PB (1.5 PB)
1.1 PB (1.5 PB)


Although there's one formula you can use to calculate the endurance of enterprise- and consumer-oriented SSDs, both classes employ different specifications. After all of the available P/E cycles are consumed, NAND cells on a consumer SSD (like the 320) retain data for 12 months. On enterprise-oriented SSDs (like the 710), it's only possible to retrieve data for three months, which is perfectly fine. In the world of enterprise storage, swapping out a defective drive occurs within hours, so a lengthy data retention window is unnecessary.

Application
Active Use
(Power On)
Data Retention
(Power Off)
Functional Failure
Requirement (FFR)
Uncorrectable Bit
Error Rate
Client
40oC
8 hrs/day
30oC
1 Year
≤ 3%
≤ 10-15
Enterprise
55oC
  24 hrs/day
30oC
  3 Months
≤ 3%
≤ 10-16
3. HET MLC: Supercharged MLC Or SLC Lite?

SSD 320: MLCSSD 320: MLCHET MLCHET MLC

It's time to pick apart High Endurance Technology MLC, because the terminology is fairly new and, frankly, subject to some confusion. According to Intel, HET provides SLC-like write endurance. This is accomplished in two ways:

  1. Die-screening consumer MLC for marginally higher endurance.
  2. Increasing the page programming cycle (tProg).

Incidentally, these two traits are what define eMLC. In other words, HET is nothing but a marketing term. At the technology level, Micron tells us that die-screening (picking out the very best dies from a wafer) optimistically results in a two-fold write endurance boost. However, the company is citing an eMLC endurance spec that's six times higher than its consumer-grade MLC. Increasing the page programming cycle makes up the difference, as it potentially increases endurance by a factor of two or three.

3x nm Lithography
SLC
MLC
eMLC
Bits/Cell
1
2
2
Endurance (P/E Cycles)
100 000
5 000
10 000 - 30 000
ECC
8b/512B
24b/1KB
24b/1KB
tProg
0.5 ms
1.2 ms
2 - 2.5 ms
tErase
1.5 - 2 ms
3 ms
3 - 5 ms
Performance Over Time
Constant
Degrades
Degrades


Although die-screening seems like an easy (albeit cost-adding) way to cherry-pick the best pieces of memory without compromising performance, increasing the time it takes to program a page doesn't necessarily sound as sexy. The reason why relates back to the difference between MLC and SLC NAND.

Single-level cell flash stores one bit per cell. It is a single-bit binary system, either a "0" or a "1." MLC memory stores up to two bits per cell, so you're looking at four states to represent all possible combinations. Though that works out neatly on paper, there is a cost associated with increasing storage density.

Flash memory only has so much voltage tolerance. You can't just double the voltage to multiply the scale. Instead, you need more sensitivity between each state. This means more programming to manipulate a very precise amount of charge stored in the floating gate. MLC and SLC memory both operate similarly. However, MLC needs more precision in charge placement and charge sensing.

As P/E cycles are slowly consumed over the course of time, however, the read margins that determine the value of each cell start shrinking as a result of:

  • loss of charge due to flash cell oxide degradation
  • over-programming caused by erratic programming steps
  • programming of adjacent erased cells due to heavy read or writes


Consequently, over time, the drive experiences data retention problems and read-related errors. Basically, they wear out. That's not a problem on SLC-based drives because they only have one reference point. But MLC memory is completely different, which is why extending the page programming cycle has a pronounced impact on endurance.

In essence, additional time is spent sending a more precise charge to the memory cell. This increases the probability of writing to a cell within a smaller window, in turn creating much larger reference points and extending the amount of wear each cell can withstand. The result is higher endurance at the expense of less performance. And that, ladies and gentlemen, is the long explanation for the SSD 710's low 2700 IOPS random write rate.

4. HET MLC: What Does Endurance Really Look Like?

Intel won't tell us exactly how many P/E cycles its 25 nm HET MLC can withstand. However, we don't have to rely solely on the company's word with regards to the 710's high endurance spec, because we can backwards-calculate the number using S.M.A.R.T. values found on Intel's latest SSDs.

Intel S.M.A.R.T.
Workload Counters
Purpose
E2
Percentage of Media Wear-out Indicator (MWI) used
E3
Percentage of workload that is read operations
E4Time counter in minutes


The media wear-out indicator is a S.M.A.R.T. value (E9) on all SSDs that tells you how many P/E cycles are used, on a scale from 100 to 1. This is like the odometer on a car. However, using this value would require months of testing, because its on a scale from 100 to 1. 

In comparison, Intel's workload counters are kind of like trip counters on a car, because they measure endurance over a fixed time period. Better yet, they provide more granular information on wear-out out, which makes it easy to measure endurance in a under a day. However, none of these workload counters are generated until the drive has been used 60 minutes or more. In practice, one hour isn't long enough for us to take a precise measurement, which is why we our endurance values are based on a 6 hour workload.

The counter starts the minute you plug in the drive, so you'll need to reset it if you want to attempt this test on your own. This is possible by sending a 0x40 instruction via smartctl.

If you're using a disk information program like CrystalDiskInfo, all S.M.A.R.T. values are in hexadecimal, which means you'll need to convert to decimal before proceeding. The E2 field is particularly unique because it's only valid out to three decimal places, and it's stored in an IEC binary format. So, after converting the E2's raw value to decimal, you have to divide by 1024 to get percentage.

Before we get to the results of our tests, we need to cover a little bit of math. If we toss out the JEDEC formula for a second, what do we know about write endurance? Rules that apply to all SSDs:

  • Host Writes ÷ NAND Writes = P/E Cycles Consumed ÷ Total P/E Cycles
  • P/E Cycles Used ÷ P/E Cycles Total = Media Wear Indicator (scale of 100 to 1)
  • 100% sequential write means Host Writes = NAND Writes (write amplification = 1)


If we take these three formulas, it's possible to calculate the write endurance of the SSD 710 using the SSD 320 as a reference point.

128 KB 100% Sequential Write
6 Hours
Intel SSD 710
200 GB
Intel SSD 320
300 GB
Total Data Written
3.88 TB
3.9 TB
Percent MWI used (E2)
0.053
0.238
Endurance In Years
1.292
0.287
Percent MWI per TB
1.35 x 10-2
6.10 x 10-2
P/E Cycles Per TB
3.07
13.7
P/E Cycles
22 337
5000
Recalculated Endurance Rating
(P/E Cycles ÷ P/E Cycles Per TB)
7268 Terabytes Written364 Terabytes Written


Starting with a 100% sequential write (write amplification equals one), we see the SSD 710's write endurance is roughly 4x to 5x higher than the SSD 320. We'll keep things simple and average out to 4.5x.

Previously, we've heard Intel mention that the NAND in its SSD 320 is rated for 5000 cycles. So, that puts the SSD 710 somewhere between 20 000 to 25 000 P/E cycles, which is in-line with what the company's competitors say eMLC should be able to do.

Now that we know what MWI looks like with a 100% sequential write, we can check write amplification in a random write workload with a high queue depth.

4 KB 100% Random Write
QD= 32, 6 Hours
Intel SSD 710
200 GB
Intel SSD 320
300 GB
Total Data Written
0.23 TB
0.11 TB
Percent MWI used (E2)
0.016
0.084
Endurance In Years
4.28
0.83
Percent MWI per TB
1.35 x 10^-2
6.10 x 10^-2
P/E Cycles Per TB
15.65
37.73
Recalculated Endurance Rating
(P/E Cycles ÷ P/E Cycles Per TB)
1437 Terabytes Written
132 Terabytes Written
Write Amplification
5.09
2.75


Interestingly, write amplification is higher on the SSD 710. However, in the same period, the 710 can write twice as much data as Intel's 320. That'd purportedly be counter to the reason for more over-provisioning, but it'll all fall into place shortly.

Perhaps more important, both drives have endurance values better than what Intel cites, which just goes to show that the JEDEC spec tends to underestimate real-world endurance. With the same random workload, all of the SSD 320's P/E cycles would be consumed in less than a year, whereas the SSD 710 could continue working for another three years or more.

5. Test Setup And Firmware Notes
Test Hardware
Processor
Intel Core i5-2500K (Sandy Bridge), 32 nm, 3.3 GHz, LGA 1155, 6 MB Shared L3, Turbo Boost Enabled
Motherboard
ASRock Z68 Extreme4, BIOS v1.4
Memory
Kingston Hyper-X 8 GB (2 x 4 GB) DDR3-1333 @ DDR3-1333, 1.5 V
System Drive
OCZ Vertex 3 240 GB SATA 6Gb/s
Tested DrivesCrucial m4 64 GB SATA 6Gb/s, Firmware: 0001

Intel SSD 510 250 GB SATA 6Gb/s, Firmware: 1.7

Intel SSD 320 300 GB SATA 3Gb/s, Firmware: 1.92

Crucial m4 128 GB SATA 6Gb/s, Firmware: 0001

Crucial m4 256 GB SATA 6Gb/s, Firmware: 0002

Crucial m4 512 GB SATA 6Gb/s, Firmware: 0001

Crucial RealSSD 256 GB SATA 6Gb/s, Firmware: 0006

OCZ Vertex 3 240 GB SATA 6Gb/s, Firmware: 2.06

OCZ Vertex 3 120 GB SATA 6Gb/s, Firmware: 2.06

OCZ Agility 3 120 GB SATA 6Gb/s, Firmware: 2.06

OCZ Solid 3 120 GB SATA 6Gb/s, Firmware: 2.06

Corsair Force 3 120 GB SATA 6Gb/s, Firmware: 1.2

Corsair Force 120 GB SATA 3Gb/s, Firmware: 2.0

Adata S511 120 GB SATA 6Gb/s, Firmware: 311A

Mushkin Chronos Deluxe 120 GB SATA 6Gb/s, Firmware: 319A

Patriot Wildfire 120 GB SATA 6Gb/s, Firmware: 319A

Kingston SSDNow V+100 128 GB SATA 3Gb/s, Firmware: CJRA

Western Digital VelociRaptor 300 GB (WD3000HLFS) SATA 3Gb/s

G.Skill FM-25S2S 64 GB SATA 3Gb/s, Firmware: 02.1

Seagate Momentus 5400.6 500 GB SATA 3Gb/s

Intel X25-M G2 160 GB SATA 3Gb/s, Firmware: 1.7

Samsung 470 256 GB SATA 3Gb/s, Firmware: AXMO

Samsung 830 256 GB SATA 6Gb/s, Firmware: CXMO

OCZ Vertex 2 (32nm) 120 GB SATA 3Gb/s, Firmware: 1.32

Kingston HyperX 240 GB SATA 6Gb/s, Firmware: 320A

Intel SSD 710 200 GB SATA 3Gb/s

Micron RealSSD P300 200 GB SATA 6Gb/s, Firmware: 0001

Corsair Force GT 240 GB SATA 6Gb/s, Firmware: 1.3

Kingston SSDNow V100 128 GB SATA 3Gb/s, Firmware: D110
Graphics
Palit GeForce GTX 460 1 GB
Power Supply
Seasonic 760 W, 80 PLUS
System Software and Drivers
Operating System
Windows 7 Ultimate 64-bit
DirectX
DirectX 11
DriverGraphics: Nvidia 270.61
RST: 10.5.0.1022
Virtu: 1.1.101
Benchmarks
Tom's Hardware Storage Bench v1.0
Trace-Based
Iometer 1.1.0
# Workers = # Logical CPUs, 4 KB Random: LBA=16 GB, varying QDs, 128 KB Sequential: QD=1
ATTO Benchmark

LBA=2 GB, QD=2 & 4, varying transfer sizes

PCMark 7
Storage Suite
Enterprise Testing: Iometer Workloads
Read
Random
Block Size
Workers
Database
67%
100%
8 KB - 100%
4
File server
80%
100%

512 Bytes – 10%

1 KB – 5%

2 KB – 5%

4 KB – 60%

8 KB – 2%

16 KB – 4%

32 KB – 4%

64 KB – 10%
4
Web server
100%
100%

512 Bytes – 22%

1 KB – 15%

2 KB – 8%

4 KB – 23%

8 KB – 15%

16 KB – 2%

32 KB - 6%

64 KB – 7%

128 KB – 1%

512 KB – 1%
4
6. Benchmark Results: Storage Bench v1.0 & PCMark 7

While Intel's SSD 710 isn't a consumer drive, PCMark 7 and our own Storage Bench v1.0 provide a quick and dirty way to examine storage performance. If you're unfamiliar with Storage Bench v1.0, we'd suggest that you read page three and four of Second-Gen SandForce: Seven 120 GB SSDs Rounded Up.

The performance specs provided for the SSD 320 and 710 are remarkably close, which is why it isn't too surprising to see similar rankings. The 710 only falls behind the 320 by a small margin, which we'd attribute to the lower random write spec inherent to eMLC NAND. However, compared to Micron's SLC-based P300, Intel's latest enterprise SSD falls far behind.

While we have enthusiast-oriented drives in this chart for comparison, the SSD 710 isn't an enthusiast product. We get good perspective on how vendors balance performance and reliability, though, in adding drives that are faster, less expensive, but ultimately unsuited for enterprise duty to the results list.

7. Benchmark Results: 4 KB Random And 128 KB Sequential Performance

An overall metric is informational, but it doesn't give us enough data about a drive's specific performance characteristics. That's why it's still important to examine random reads, random writes, sequential reads, and sequential writes. If you're unfamiliar with those terms and what they represent, you can refer back to page three of SSD Performance In Crysis 2, World Of Warcraft, And Civilization V.

Even though the SSD 710 is an enterprise-class drive, suggesting that it's intended to handle intensive workloads where I/O operations stack up, we're starting our synthetic testing with a queue depth of one in order to compare baseline performance with other SSDs. We'll get to the enterprise workloads shortly.

The random read rate for the 710 is remarkably similar to the 320. But again, that's hardly a surprise.

The SSD 710 has a specified random write speed nearly 10x lower than the 320, but that only applies to the high queue depths where all SSD vendors benchmark their offerings, yielding the highest possible performance. When we scale back to a queue depth of one, the SSD 710 behaves a lot like Intel's SSD 320 yet again.

Intel's datasheet also suggests nearly identical sequential performance between the enterprise and client products. Again, though, this only applies to higher queue depths. With only one active I/O operation, the SSD 710 performs about 15% better in sequential reads, and similarly in sequential writes.

Compared to the P300, the SSD 710 offers ~50% slower performance for both sequential reads and writes. You have to ask yourself it the Micron drive is worth its price premium, though; at 200 GB, it sells online for a little less than $2000.

8. Benchmark Results: Enterprise Performance

You'll see periods of low activity in any type of workload, even in enterprise applications. However, more so in data centers than desktop workstations it's safe to expect random access with a large number of outstanding I/O operations.

In random reads, the SSD 710 offers similar performance to the 320, even at higher queue depths.

Writes are another story, though. Once we move to a queue depth of four (effectively 16, since we're using four workers), the SSD 710 starts to pull ahead of the 320. But, at best, it's only able to achieve a 20-30% lead. Obviously, that's not enough to match the speed of SATA 6Gb/s SSDs, which makes sense since Intel's proprietary controller is a 3 Gb/s device.

Why does the SSD 710 appear to outperform the SSD 320 when its write spec is supposed to be less aggressive? In enterprise-class applications, the system always writes random data to the entire LBA space, since the SSD has a tendency to be empty (consider the drive used as a cache). As the LBA space increases, random write performance drops because the controller needs to perform more disk operations, such as garbage collection, to maintain health and performance. Conversely, consumer drives are at least partially filled with persistent data, which is why performance is measured in a fixed LBA space. However, we're testing relative performance, so the LBA test space is fixed to just 16 GB.

Subjected to a database workload, the SSD 710 offers better performance than its consumer-oriented counterpart, but speeds are still relatively (and understandably) poor compared to competing SATA 6Gb/s-based SSDs. When we stripe two 710s, performance improves, though not linearly. At best, performance in RAID 0 still falls short of SLC-based SSDs like the P300.

The file server profile employs a more read-heavy workload, which is why the 710 scales better in RAID. However, this is also an area where most SATA 6Gb/s SSDs really shine.

This is particularly evident at the extreme end, where the Vertex 3 hits ~33 000 IOPS. That's 5000 IOPS higher than two 710s in RAID. Though the Vertex 3 isn't directly comparable here, the Vertex 3 Pro would be, and it performs very similarly. That drive still has to prove itself in an enterprise environment, though. When we surveyed data centers for Investigation: Is Your SSD More Reliable Than A Hard Drive?, all of them were running Intel drives. Intel is unquestionably the incumbent in a competition that puts reliability on a higher pillar than raw speed.

The Web server profile is similar to the previous file server workload. It's composed of 100% reads and gives more weight to smaller transfer sizes. As a result, we finally see a case where a pair of 710s in RAID 0 can surpass the SLC-based P300 and speedy Vertex 3. However, this only occurs at queue depths higher than 16. In a single-drive configuration, the SSD 710 only leads the 320 by a small margin.

9. Sequential Performance Versus Transfer Size

There are a number of enterprise-oriented workloads that push a lot of random data. Exchange information stores and online transaction processing are both good examples. Sequential tasks might include writing out log files or backing up data to disk. Either way, though, you're going to see more outstanding I/O operations using those applications than in a desktop environment, and that's why we're configuring ATTO to use its maximum queue depth of 10.

At high queue depths, RAID really delivers better performance at smaller transfer sizes. There is a performance overhead incurred with every page request (8 KB), so scaling up using striping is the only way to improve our results in that discipline.

While the RAID array stands apart in read operations, it's not as dominant in our write test; the SSD 710s deliver about ~40% more performance. In a single-drive configuration, the 710 continues to perform like the 320.

SandForce's compression technology really puts things into perspective, as the Vertex 3 leaves other SSDs in the dust at transfer sizes greater than 32 KB. It's the only SSD of the group to easily break through the 500 MB/s barrier, though it relies on the fact that ATTO uses compressible data (like log files) to achieve those numbers. Subject the same drive to incompressible data at the same queue depth and throughput drops to about 240 MB/s (using 128 KB blocks) instead.

10. Performance Over Time

According to the slide above, which we took from this year's Flash Memory Summit in Santa Clara, there are a handful of assumptions made about enterprise environments as they relate to desktops. Enterprise drives are available 24x7, they need to be evaluated after hitting steady-state performance, no down-time is accepted, and the consequences of a failure are catastrophic. 

When a drive is getting hit all day, every day, and it's operating at its steady-state point, performance needs to be both acceptable and predictable. If a server's workload is running for an extended period, it won't sit idle long enough for background garbage collection to move scattered pages into single blocks, restoring performance and reducing write amplification. Naturally, that's bad if the drive is unable to cope.

Clean Performance

Intel SSD 710: Clean PerformanceIntel SSD 710: Clean Performance

Intel SSD 320: Clean PerformanceIntel SSD 320: Clean Performance

Examining how a drive might perform over time isn't that difficult. First, we just have to fill up all user-accessible space using a sequential write, making the drive "dirty." Then, we subject it to a 4 KB random write with a queue depth of 32. Because the drive is full of data, though, garbage collection can't consolidate scattered pages into free blocks. When we start writing sequential data again, the effects of active garbage collection kick in.

Random Writes, 20 Min.

Intel SSD 710Intel SSD 710

Intel SSD 320Intel SSD 320

If the drive recovers quickly, you can be fairly certain that there's lots of active garbage collection going on.

When we subject the SSD 710 to 20 minutes of torturous random writes, we start to see small differences. Whereas the 320 employs foreground garbage collection over an extended period, the 710 has a tendency to perform a lot of garbage collection all at once. As a result, we only see one dip in performance, recovered from relatively quickly.

That's not the only difference, though. When you look at the 320's chart, it's apparent that some garbage collection even occurs during read operations. We get confirmation when we revisit the endurance test. After running the database profile for six hours, we see a higher write amplification value.

The SSD 710 doesn't do any garbage collection on reads, but its write amplification goes down as a result of the 40% over-provisioning, which decreases the amount of data rearrangement necessary to optimize performance.

Endurance Calculations
(Workload Counters)
Intel SSD 710
200 GB
Intel SSD 320
300 GB
4 KB 100% Random Write
QD= 32, 6 Hours
WA = 5.09
1437 TBW
WA = 2.75
132 TBW
Database 67% Random Reads
QD =32, 6 Hours
WA = 4.03
1818 TBW
WA  = 3.49
104 TBW


Perhaps all of those pretty pictures depicting performance up in the 100 MB/s range paint an overly optimistic picture of performance, though. Hammering the SSD 320 with 4 KB writes for 20 minutes still represents a fairly desktop-oriented workload. If we sustain that workload for hours, as you might see in an enterprise, random writes fall to as low as 20 MB/s. When we subject Intel's SSD 710 to an hour of random writes, its advantage over the desktop-oriented hardware becomes more clear.

Random Writes, 60 Min.

Intel SSD 710: 60 Minutes Random WriteIntel SSD 710: 60 Minutes Random Write            

Intel SSD 320: 60 Minutes of Random WriteIntel SSD 320: 60 Minutes of Random Write

According to Iometer, sequential read/write performance should be in the 175-200 MB/s range. Performance drops precipitously, though, as very little garbage collection occurs in real-time.

If we combine these results with our endurance test, we see that the 710 handles foreground garbage collection more adeptly, thanks in part to the large amount of over-provisioning. As a whole, this contributes to a minimum sequential write speed of 60 MB/s. In comparison, the 320 relies more on background garbage collection (particularly during reads) in order to recover performance.

After 30 Min. Idle

Intel SSD 710: After 30 Minutes of IdleIntel SSD 710: After 30 Minutes of Idle

Intel SSD 320: After 30 Minutes of IdleIntel SSD 320: After 30 Minutes of Idle

In either case, if you give the drives some idle time, performance recovers to a clean state, even without TRIM.

11. Intel's SSD 710: Making Enterprise Storage More Affordable?

It's almost ironic that the enterprise segment, which can often put the highest-performance hardware to use most immediately, also has to be the most cautious with unproven technology. It has taken a long, long time for solid-state storage to earn its place in data centers, but now SSDs are smashing bottlenecks in the server world, just as they did on the desktop more than three years ago when Intel's X25-M first pushed them into the mainstream.

Back then, the enterprise-oriented X25-Es were putting up the most impressive numbers. However, they were both comparatively small and expensive, affecting their accessibility. Now, the SSD 710 presents us with what we must presume to be a much more mature product based on a controller that has evolved over a very long life and HET MLC memory claimed to outlast the compute-quality NAND used in desktop drives. At the same time, we're forced to accept a compromise. In the interest of pulling down prices and pushing higher-capacity models, performance gets de-emphasized through a 3 Gb/s controller.

Are enterprises willing to accept such a trade? That's hard for us to say. On one hand, our own research suggests that Intel sets the standard for SSD reliability. And a business previously limited to short-stroked hard drives might jump at the opportunity to pick up high-capacity solid-state storage for a lot less per gigabyte than the now-defunct X25-E. On the other, there are a number of applications that absolutely need as much throughput as possible, hence the growing popularity of PCIe-based SSD unconstrained by SATA.

Oh, but guess what? Intel has plans to address that segment, too, with its upcoming SSD 720, armed with SLC NAND and a PCI Express interface. More on that later.

Intel SSD 320 & SSD 710Intel SSD 320 & SSD 710

What we can say for the SSD 720 is that Intel's HET MLC-based drive really should offer exceptional write endurance. Going by the numbers, we calculated a 1818 terabyte-written value for our 200 GB SSD 710. That's 17x times higher than the consumer-oriented SSD 320. It's not the 33x margin Intel suggested at IDF, but we can certainly attest to this drive's enterprise pedigree. 

Initially, the company's warranty policy on the SSD 710 did worry us because it included some fairly non-standard verbiage. Most of Intel's SSDs have a flat three-year warranty (except for the SSD 320, which is five years), but the terms for the 710 are three years or when the media wear-out indicator (E9) reaches 1, whichever comes first. With a bit of clever math, however, we found that it would take 4.2 years to consume all of the 200 GB SSD 710's P/E cycles, assuming a 100% 4 KB random write workload, 24x7, at a queue depth of 32. That's about 880 GB per day of data, by the way. Compare that to a 300 GB SSD 320, which would be worn out in a year.

Introduce a bit of idle time and make two-thirds of the drive accesses reads and Intel's SSD 710 looks like it has more than six or seven years in it before potentially wearing out. And again, that's 24x7 activity. Surprisingly, sequential transfers turn out to be the enemy here. While write amplification is low, data moves at a pretty zippy 200 MB/s or so, moving up to 15.5 TB a day and potentially wearing the drive out significantly faster.  

And so, with the SSD 710's enterprise chops fully explored, we settle in to the debate over cost. There's nothing like Intel's new drive on the market, so have to compare a 200 GB SSD 710 at $1200 to a 200 GB P300 at $2100. Intel is commendably introducing cost-conscious businesses to a relatively affordable eMLC-based product compared to its outgoing SLC-based solution. To that end, the SSD makes sense if you need lots of capacity and enterprise-class endurance. Performance-sensitive applications are still best-served by SLC-based drive like the P300. Intel plans to address its position in the that particular market soon, though, with its SSD 720.