Sign in with
Sign up | Sign in
Samsung 850 Pro SSD Review: 3D Vertical NAND Hits Desktop Storage
By ,
1. Samsung 850 Pro SSD: Introducing V-NAND

Update: Samsung is cutting prices on the 850 Pro prior to availability later this month. While the 128 GB remains unchanged, the three larger capacities will each sell for $30 less than Samsung's initial MSRP.

I liked Samsung's 840 EVO. Although it was hobbled in some ways by triple-level-cell NAND, clever features and technologies helped propel the drive to unlikely greatness. Whereas the original 840 suffered from slow write speeds, the EVO broke through with an emulated SLC-like cache, bridging the considerable performance gap between two- and three-bit-per-cell flash.

Tack on the host-side caching software (originally sourced from Nvelo) through Samsung's Magician software and the EVO flew even faster. Not that it needed much help; the EVO's TurboWrite system helped propel the 840's successor into some fairly elite company. Not bad for a supposedly mainstream SSD. Once Samsung promised me that the EVO would eventually sport the same Crucial M500-class encryption features through a firmware update, I made the call to award it my first ever (and only) Tom's Hardware Smart Buy recognition. It's the SSD I use in my personal machine, and it's also something I recommend to friends.

That's not to say I'm a huge fan of three-bit-per-cell flash, with its lower endurance, higher latency, and error-prone nature. But it certainly appears that Samsung figured how to mitigate a lot of the technology's inherent disadvantages in the EVO's case. Combining an in-house controller, custom firmware, and the company's own flash proved to be a stout combination. So much so, in fact, that I'll admit liking the 840 EVO more than the 840 Pro in most desktop applications, particularly as you factor in price. The old Pro just doesn't do much for me.

So where does that put Samsung's new 850 Pro in the hierarchy of storage?

For starters, it retains the company's MEX controller, used also in the 840 EVO. Operating at 400 MHz, it offers a considerable amount of processing power. And like other storage controllers, it employs multiple purpose-built execution cores. There's a lot going on in modern SSDs, so we've seen the hardware inside become increasingly potent. The older 840 Pro's silicon operates 100 MHz slower, and doesn't support Samsung's newest features. Still, I believe that drive's controller hardware is still similar.

First generantion V-NAND used 24 layers.First generantion V-NAND used 24 layers.

The most significant differences involve Samsung's new 3D V-NAND, which represents a radical change in the way flash is designed and constructed. Last year, at the 2013 Flash Memory Summit, I attended a keynote given by Dr. E.S. Jung, executive vice president, semiconductor R&D center at Samsung. In front of a packed crowd, he raised the curtain on the company's next generation of flash technology. It's rare that a speech brings down the house at an event like FMS, but the advances promised by V-NAND have wide-reaching implications, so the reaction was understandable.

In short, we've been hearing for a while now that NAND can't advance much further (just as we've heard Moore's Law is at its end). Invariably, some technology comes along to help circumvent the limitations standing in front of progress. As it happens, NAND cells become more error-prone as feature size shrinks. The number of program/erase cycles they can withstand drops precipitously, and operations to one location in a planar array can cause unintended changes in adjacent cells.

V-NAND sees 32 cells stacked vertically in cylindrical "rods". Whereas planar NAND puts the cells adjacent together in lines, then packs multiple lines together to create one die, imagine V-NAND as stacks of Pringles cans. Where two cans join vertically, you have a word line. Connect each columnar stack of potato chips together like telephone poles, and you get the bit lines. In this way, it's still NAND as we know it. But stacking the columns higher reduces the cell-to-cell interference encountered in planar NAND.

This is how Samsung aims to continue shrinking feature size, though purportedly, V-NAND is also faster and more power efficient as well. It helps make better use of package real estate, too. Build up instead of out, and you squeeze more stuff into the same amount of floor space. That's why closets get shelves. This technology was years in the making, and now it's the centerpiece of Samsung's 850 Pro.

So that's really what this new SSD is all about. It sports the same (proven) 400 MHz MEX controller lashed to a bunch of V-NAND. With its upgrades, the 850 Pro should serve up I/O faster whilst using less power. It's not supposed to be radically quicker than the 840 Pro, but then again, it's still limited to aged SATA 6Gb/s and AHCI standards. So, there's really not much room for it go faster. I don't want to give too much of my testing away, but the real steps forward involve service times, quality, and latency. In this case, the spec sheet doesn't tell the whole story. But then, it never does...

The specifications tell us that random 4 KB IOPS at a queue depth of one are in the same ballpark as the 840 EVO. This number is almost wholly a function of NAND interface speed, so it appears Samsung's V-NAND operates at the same performance level as the EVO's three-bit-per-cell flash. That's good enough for a small bump compared to the 840 Pro. 

Samsung is reluctant to provide details about its NAND, so it's difficult to comment on certain architectural features like die count. Regardless, the 850 Pro gets 10 years of warranty coverage, mirroring recent developments from SanDisk.

Lastly, Samsung claims the 850 Pro drops as low as 2 mW in DevSlp mode. We have the equipment to verify this, and we'll run the numbers. Active idle falls in the .4 W range, according to the specs. Neither figure surprises me, given this company's recent history with high efficiency.

The next step: open these things up to see what they look like inside. I love this part.

2. Inside Of Samsung's 850 Pro

We're dealing with the same chassis as the original 840. It's far better than the old 830 when it comes to accessing the internals, but you do need pentalobe drivers to remove three screws. Do I even need to mention that doing this defaces the pristine label and voids your warranty? Probably not.

Also like the 840 EVO, PCB size depends on capacity. Of the three models I'm testing, none fill the entire enclosure. Instead, we see this:

The 1024 GB drive (left) employs a slightly longer board, with four packages on the front and four on the back. The 128 GB version (right) gets just four packages total.

Here's a close-up:

1024 GB 850 Pro PCB1024 GB 850 Pro PCB

Isn't that MEX controller pretty? It wears 1 GiB worth of LPDDR2 like a funny hat in this 1024 GB implementation. Samsung's Pro family includes 1 MB of cache per GB of capacity, so the 128 and 256 GB drives come with 128 and 256 MB worth of DRAM.

The NAND part numbers aren't decipherable without an updated decoder. For now, I can tell you that the packages you see on-board all come from Samsung. I'm making an effort to dig up additional details, but it'll probably take a trip to South Korea for answers. Of course, if you're reading this on launch day, that's exactly where I'll be, too.

3. How We Tested Samsung's 850 Pro

Our consumer storage test bench is based on Intel's Z77 Platform Controller Hub paired with an Intel Core i5-2400 CPU. Intel's 6- and 7-series chipsets are virtually identical from a storage perspective. We're standardizing on older RST 10.6.1002 drivers for the foreseeable future.

Updates to the RST driver package occasionally result in subtle performance changes. They can also lead to some truly profound variance in scores and results as well, depending on the revision. Some versions flush writes more or less frequently. Others work better in RAID situations. Builds 11.2 and newer support TRIM in RAID as well. Regardless, results obtained with one revision may or may not be comparable to results obtained with another, so sticking with one version across all testing is mandatory.

Test Hardware
ProcessorIntel Core i5-2400 (Sandy Bridge), 32 nm, 3.1 GHz, LGA 1155, 6 MB Shared L3, Turbo Boost Enabled
MotherboardGigabyte G1.Sniper M3
MemoryG.Skill Ripjaws 8 GB (2 x 4 GB) DDR3-1866 @ DDR3-1333, 1.5 V
System Drive Intel S3500 480 GB SATA 6 Gb/s, Firmware: 0306
Drive(s) Under TestSamsung 850 Pro 128 GB SATA 6 Gb/s, Firmware: EXM01B6Q

Samsung 850 Pro 256 GB SATA 6 Gb/s, Firmware: EXM01B6Q

Samsung 850 Pro 1024 GB SATA 6 Gb/s, Firmware: EXM01B6Q
Comparison DrivesPlextor M6e 256 GB M.2 PCIe x2, Firmware: 1.00

Plextor M6S 256 GB SATA 6 Gb/s, Firmware: 1.00

Plextor M6M 256 GB mSATA 6 Gb/s, Firmware: 1.00

Adata SP920 1024 GB SATA 6 Gb/s, Firmware: MU01

Adata SP920 512GB SATA 6 Gb/s, Firmware: MU01

Adata SP920 256 GB SATA 6 Gb/s, Firmware: MU01

Adata SP920 128 GB SATA 6 Gb/s, Firmware: MU01

Crucial M550 1024 GB SATA 6 Gb/s, Firmware: MU01

Crucial M550 512 GB SATA 6 Gb/s, Firmware: MU01

Intel SSD 730 480 GB SATA 6 Gb/s, Firmware: L2010400

Samsung 840 EVO mSATA 120 GB, Firmware: EXT41B6Q

Samsung 840 EVO mSATA 250 GB, Firmware: EXT41B6Q

Samsung 840 EVO mSATA 500 GB, Firmware: EXT41B6Q

Samsung 840 EVO mSATA 1000 GB, Firmware: EXT41B6Q

SanDisk X210 256 GB, Firmware X210400

SanDisk X210 512 GB, Firmware X210400

Intel SSD 530 180 GB SATA 6Gb/s, Firmware: DC12

Intel SSD 520 180 GB SATA 6Gb/s, Firmware: 400i

Intel SSD 525 180 GB mSATA, Firmware: LLKi

SanDisk A110 256 GB M.2 PCIe x2, Firmware: A200100

Silicon Motion SM226EN 128 GB SATA 6Gb/s, Firmware: M0709A

Crucial M500 120 GB SATA 6Gb/s, Firmware: MU02

Crucial M500 240 GB SATA 6Gb/s, Firmware: MU02

Crucial M500 480 GB SATA 6Gb/s, Firmware: MU02

Crucial M500 960 GB SATA 6Gb/s, Firmware: MU02

Samsung 840 EVO 120 GB SATA 6Gb/s, Firmware: EXT0AB0Q

Samsung 840 EVO 240 GB SATA 6Gb/s, Firmware: EXT0AB0Q

Samsung 840 EVO 480 GB SATA 6Gb/s, Firmware: EXT0AB0Q

Samsung 840 EVO 1 TB SATA 6Gb/s, Firmware: EXT0AB0Q

SanDisk Ultra Plus 64 GB SATA 6Gb/s, Firmware: X211200

SanDisk Ultra Plus 128 GB SATA 6Gb/s, Firmware X211200

SanDisk Ultra Plus 256 GB SATA 6Gb/s, Firmware X211200

Samsung 840 Pro 256 GB SATA 6Gb/s, Firmware DXM04B0Q

Samsung 840 Pro 128 GB SATA 6Gb/s, Firmware DXM04B0Q

SanDisk Extreme II 120 GB, Firmware: R1311

SanDisk Extreme II 240 GB, Firmware: R1311

SanDisk Extreme II 480 GB, Firmware: R1311

Seagate 600 SSD 240 GB SATA 6Gb/s, Firmware: B660

Intel SSD 525 30 GB mSATA 6Gb/s, Firmware LLKi

Intel SSD 525 60 GB mSATA 6Gb/s, Firmware LLKi

Intel SSD 525 120 GB mSATA 6Gb/s, Firmware LLKi

Intel SSD 525 180 GB mSATA 6Gb/s, Firmware LLKi

Intel SSD 525 240 GB mSATA 6Gb/s, Firmware LLKi

Intel SSD 335 240 GB SATA 6Gb/s, Firmware: 335s

Intel SSD 510 250 GB SATA 6Gb/s, Firmware: PWG2

OCZ Vertex 3.20 240 GB SATA 6Gb/s, Firmware: 2.25

OCZ Vector 256 GB SATA 6Gb/s, Firmware: 2.0

Samsung 830 512 GB SATA 6Gb/s, Firmware: CXMO3B1Q

Crucial m4 256 GB SATA 6Gb/s Firmware: 000F

Plextor M5 Pro 256 GB SATA 6Gb/s Firmware: 1.02

 Corsair Neutron GTX 240 GB SATA 6Gb/s, Firmware: M206
Graphics
MSI Cyclone GTX 460 1 GB
Power Supply
Seasonic X-650, 650 W 80 PLUS Gold
Chassis
Lian Li Pitstop T60
RAID
LSI 9266-8i PCIe x8, FastPath and CacheCade AFK
System Software and Drivers
Operating
System
Windows 7 x64 Ultimate
DirectX
DirectX 11
Drivers
Graphics: Nvidia 314.07
RST: 10.6.1002
IMEI: 7.1.21.1124
Generic AHCI: MSAHCI.SYS
Benchmarks
ULINK DriveMaster 2012
DM2012 v980, JEDEC 218A-based TRIM Test, Protocol Test Suite
Test Specific Hardware
SAS/SATA Power Hub, DevSlp Platform
Tom's Hardware Storage Bench v1.0
Intel iPeak Storage Toolkit 5.2.1, Tom's Storage Bench 1.0 Trace Recording
Iometer 1.1.0# Workers = 1, 4 KB Random: LBA=16 GB, varying QDs, 128 KB Sequential, 16 GB LBA Precondition, Exponential QD Scaling
PCMark 8
PCMark 8 2.0.228, Storage Consistency Test
PCMark 7
Secondary Storage Suite
4. Results: 128 KB Sequential Read And Write

Fantastic sequential read and write performance is a trademark of modern SSDs. To measure it, we use incompressible data over a 16 GB LBA space, and then test at queue depths from one to 16. We're reporting these numbers in binary (where 1 KB equals 1024) instead of decimal numbers (where 1 KB is 1000 bytes). When necessary, we also limit the scale of the chart to enhance readability.

128 KB Sequential Read

There's not much difference between drives in this metric. Then again, I didn't expect any. Getting 538 MiB/s is about all we'll see from SATA 6Gb/s, and these could have been any three drives to pass through my lab lately; they're all hitting the interface's ceiling. On to the writes.

128 KB Sequential Write

The 128 GB 850 Pro is the only drive to register lower performance, as we knew would be the case. But hitting 460 MB/s is no joke. No other 128 GB drive I have comes close. Note that, first, I'm using 128 KB access sizes, and second, reporting in binary, not decimal. In decimal, that'd be 482 MB/s. Put another way, the 128 GB 850 Pro pushes 100 MB/s more than the 840 Pro at 128 GB.

Samsung's 256 and 1024 GB models are blazing fast, demonstrating throughput in excess of 510 MB/s (again, in binary). That's as fast as you're going to see from a SATA 6Gb/s-based SSD.

Here's a breakdown of the maximum observed 128 KB sequential read and write performance with Iometer:

Sorted by combined maximum read and write throughput, two 850 Pros establish beachheads just below the pair of PCIe-based SSDs. No drive with a SATA interface can claim to deliver more sequential performance. And again, dig through this chart and you'll find that no competing 128 GB-class offering touches the 850 Pro.

5. Results: 4 KB Random Read And Write

We turn to Iometer as our synthetic metric of choice for testing 4 KB random performance. Technically, "random" translates to a consecutive access that occurs more than one sector away. On a mechanical hard disk, this can lead to significant latencies that hammer performance. Spinning media simply handles sequential accesses much better than random ones, since the heads don't have to be physically repositioned. With SSDs, the random/sequential access distinction is much less relevant. Data are put wherever the controller wants it, so the idea that the operating system sees one piece of information next to another is mostly just an illusion.

4 KB Random Reads

Testing the performance of SSDs often emphasizes 4 KB random reads, and for good reason. Most system accesses are both small and random. Moreover, read performance is arguably more important than writes when you're talking about typical client workloads.

If IOPS were the sole performance metric of solid-state storage, Samsung's 850 Pro would rule the market unchallenged. Of course, that's not how it works, so there plenty of benchmarks left to pore over. But, if you want to beat the 850 Pro in this discipline specifically, you have to reach up to the 512 GB Samsung XP941 on four lanes of PCI Express. As spec'ed, the 850s hit right around 10,000 IOPS at a queue depth of one, then smoothly step up to 100,000 IOPS at a queue depth of 32.

Does it matter if a desktop-oriented drive does 90,000 or 100,000 IOPS with 32 outstanding commands? I say no. Samsung could have artificially limited these things to 80,000 IOPS and they'd be just as fast in most of our metrics. But we're getting an indication of the underlying platform's potential.

I'll wrap this section by pointing out that 100,000 4 KB IOPS is 400 MB/s worth of throughput, or roughly 100x what a 15,000 RPM SAS drive offers. 

4 KB Random Writes

The 850 Pros peak just north of 90,000 IOPS with eight outstanding commands. It isn't common to see three drives from 128 to 1024 GB nail the same random performance. Typically, the smaller capacities underperform larger drives. Newer models tend to include eight dies, while 1 TB SSDs wield 64. That's a lot more dies to spread writes across in random workloads. More than likely, Samsung is arming its smaller 850 Pro with less-dense flash, increasing die count to achieve the same effect.

There's no question that Samsung's 128 GB 850 Pro is shaping up to be the fastest SSD I've ever tested in that size. Of course, you could double your capacity for the same money by buying something else. But the allure of speed is hard to ignore.

Here's a break-down of the maximum observed 4 KB sequential read and write performance with Iometer. The order the drives appear in our chart is determined by maximum combined read and write performance.

I threw in the XP941 just for the heck of it. Otherwise, we'd see a clean sweep by Samsung's 850 Pros, which even topple the mighty Plextor M6e and SanDisk A110 PCI Express-based drives. Only the XP941's crazy four-lane PCIe interface can get past the trio of 850s.

6. Results: Tom's Hardware Storage Bench v1.0

Storage Bench v1.0 (Background Info)

Our Storage Bench incorporates all of the I/O from a trace recorded over two weeks. The process of replaying this sequence to capture performance gives us a bunch of numbers that aren't really intuitive at first glance. Most idle time gets expunged, leaving only the time that each benchmarked drive is actually busy working on host commands. So, by taking the ratio of that busy time and the the amount of data exchanged during the trace, we arrive at an average data rate (in MB/s) metric we can use to compare drives.

It's not quite a perfect system. The original trace captures the TRIM command in transit, but since the trace is played on a drive without a file system, TRIM wouldn't work even if it were sent during the trace replay (which, sadly, it isn't). Still, trace testing is a great way to capture periods of actual storage activity, a great companion to synthetic testing like Iometer.

Incompressible Data and Storage Bench v1.0

Also worth noting is the fact that our trace testing pushes incompressible data through the system's buffers to the drive getting benchmarked. So, when the trace replay plays back write activity, it's writing largely incompressible data. If we run our storage bench on a SandForce-based SSD, we can monitor the SMART attributes for a bit more insight.

Mushkin Chronos Deluxe 120 GB
SMART Attributes
RAW Value Increase
#242 Host Reads (in GB)
84 GB
#241 Host Writes (in GB)
142 GB
#233 Compressed NAND Writes (in GB)
149 GB

Host reads are greatly outstripped by host writes to be sure. That's all baked into the trace. But with SandForce's inline deduplication/compression, you'd expect that the amount of information written to flash would be less than the host writes (unless the data is mostly incompressible, of course). For every 1 GB the host asked to be written, Mushkin's drive is forced to write 1.05 GB.

If our trace replay was just writing easy-to-compress zeros out of the buffer, we'd see writes to NAND as a fraction of host writes. This puts the tested drives on a more equal footing, regardless of the controller's ability to compress data on the fly.

Average Data Rate

The Storage Bench trace generates more than 140 GB worth of writes during testing. Obviously, this tends to penalize drives smaller than 180 GB and reward those with more than 256 GB of capacity.

Again, I added results from Samsung's 512 GB XP941 over two and four PCIe lanes. Also making an appearance is the 250 GB 840 EVO with RAPID enabled. Those results are in orange, helping put the 850 Pro's performance into context.

The 128 GB 850 Pro sweeps past even the PCIe-based Samsung XP941 over two second-gen PCIe lanes. The larger models are faster still. Only Samsung's XP941 communicating across four PCIe lanes bests the 850 Pros.

Service Times

Beyond the average data rate reported on the previous page, there's even more information we can collect from Tom's Hardware's Storage Bench. For instance, mean (average) service times show what responsiveness is like on an average I/O during the trace.

Write service time is simply the total time it takes an input or output operation to be issued by the host operating system, travel to the storage subsystem, commit to the storage device, and have the drive acknowledge the operation. Read service is similar. The operating system asks the storage device for data stored in a certain location, the SSD reads that information, and then it's sent to the host. Modern computers are fast and SSDs are zippy, but there's still a significant amount of latency involved in a storage transaction.

Mean Read Service Time

Whatever medieval magic that animates the 850 Pros pushes read service times into previously-unseen territory. Without more detail on this drive's tweaks, I have to assume the subtle improvement over Samsung's previous-gen offerings comes from V-NAND.

Mean Write Service Time

Crucial's 1 TB M550 splits the smaller 850 Pros. But the two larger Samsung SSDs break new ground otherwise.

7. Results: PCMark 8's Expanded Storage Testing

Futuremark's PCMark 8 expanded storage tests are awesome. With so much data and a comprehensive testing regimen, we can really drill down on drive performance.

First, the raw block device (there is no partition) is preconditioned twice by filling the entire accessible LBA space with 128 KB sequential writes. Once that is completed, the first Degradation Phase randomly writes blocks between 4 KB and 1 MB in size to random LBA spaces on the drive. Since the writes aren't 4 KB-aligned much of the time, the SSD's performance drops quickly. After all, non-4 KB-aligned accesses create overhead and generally increase write amplification significantly.

The first Degradation Phase begins with 10 minutes of those punishing random offset writes, after which each PCMark 8 activity trace is played against the SSD being tested. The successive degradation rounds are similar, except an additional five minutes are tacked onto each iteration. After eight repetitions, that write period expands to 45 minutes.

Next comes the Steady Phase. Each of five Steady Phases writes 45 minutes worth of random offset data prior to trace playback, pushing the drive even harder and making it more difficult to perform housekeeping duties. With fewer blocks available for writing, latency increases substantially.

Lastly, PCMark 8 moves into a Recovery Phase, which consists of five idle minutes before trace playback. Repeat that five times, and the test concludes.

For more information on the test and how it works, check out Plextor M6e 256 GB PCI Express SSD Review: M.2 For Your Desktop.

Storage Consistency With PCMark 8's Adobe Photoshop (Heavy) Trace

Because there are 18 individual rounds packed with 10 traces each, we need to focus. We'll choose one trace, Adobe Photoshop (Heavy), and keep tabs on it through the entire extended run.

Bandwidth

The Bandwidth chart simply reflects per-iteration throughput for my preferred Adobe Photoshop (Heavy) trace.

The 850 Pros are interesting in that they don't bottom out until they're several rounds into the testing. In comparison, the 256 GB 840 Pro looks old and sad. Even the 840 EVO fares better. During the Degrade and Steady phases, it's SanDisk's X210 that appears to hold on best, though.

As the Recovery phases progress, all three 850 Pros spike and never look back. This behavior is unusual as well, though this time in a good way (like losing your wallet and someone tracking you down to return it).

Latency

In this test, we take that same Adobe Photoshop (Heavy) trace and use average read and write latency to illustrate responsiveness. I'll sprinkle in competing drives for flavor, too.

The yellow line represents Samsung's older 840 Pro at 256 GB, and it's clearly not doing well (since low latencies are good; tall bars are not). The 128 GB 850 Pro is outmoded by Intel's SSD 730. However, Samsung's two larger capacities hang right in there.

The 128 GB 850 Pro doesn't handle this workload as deftly. But Samsung's 256 and 1024 GB models excel. Still, they fall a bit short of SanDisk's X210 until the Recovery phases hit. Frankly, I should have been more aggressive about recognizing the X210 back when I reviewed it.

Best and Worst Score Reference

The best scores of the 18 rounds are reported, along with the worst. Here they are for comparison.

8. Results: TRIM Testing With DriveMaster 2012

We've been utilizing ULINK's DriveMaster 2012 software and hardware suite to introduce a new test for client drives. Using JEDEC's standardized 218A Master Trace, DriveMaster can turn a sequence of I/O (similar to our Tom's Hardware Storage Bench) into a TRIM test. JEDEC's trace is months and months of drive activity, day-to-day activities, and background operating system tasks.

ULINK strips out the read commands for this benchmark, leaving us with the write, flush, and TRIM commands to work with. Execute the same workload with TRIM support and without, and you end up with a killer metric for further characterizing drive behavior.

DriveMaster is used by most SSD manufacturers to create test cases and perform specific measurements. It's currently the only commercial product that can create the scenarios needed to validate TCG Opal 2.0 security, though it's almost unlimited in potential applications. Much of the benefit tied to a solution like DriveMaster is its ability to diagnose bugs, ensure compatibility, and issue low-level commands. In short, it's very handy for the companies actually building SSDs. And if off-the-shelf scripts don't do it for you, make your own. There's a steep learning curve, but the C-like environment and command documentation gives you a fighting chance.

This product also gives us some new ways to explore performance. Testing the TRIM command is just the first example of how we'll be using ULINK's contribution to the Tom's Hardware benchmark suite.

On a 256 GB drive, each iteration writes close to 800 GB of data, so running the JEDEC TRIM test suite once on a 256 GB SSD generates almost 3.2 TB of mostly random writes (it's 75% random and 25% sequential). By the end of each run, over 37 million write commands are issued.

The first two tests employ DMA to access the storage, while the last two use Native Command Queuing. Since most folks don't use DMA with SSDs (aside from some legacy or industrial applications) we don't concern ourselves with those. It can take up to 96 hours to run one drive through all four runs, though faster drives can roughly cut the time in half. Because so much information is being written to an already-full SSD (the drive is filled before each test, and then close to 800 GB are written per iteration), SSDs that perform better under heavy load fare best. Without TRIM, on-the-fly garbage collection becomes a big contributor to high IOPS. With TRIM, 13% of space gets TRIM'ed, leaving more room for the controller to use for maintenance operations.

TRIM Testing

Average

To avoid adding too much data, I have the average performance for each benchmarked SSD with and without TRIM support enabled. Displayed in IOPS, this helps us make comparisons more quickly.

One terabyte is a lot of flash. All of that capacity helps the 1024 GB 850 Pro achieve excellent results with and without TRIM support (both scores are close to equal). Drop to the 256 GB model, though, and TRIM support becomes crucial for attaining peak performance. The 128 GB drive only exaggerates the situation; it needs the TRIMed space more than it needs the latency reduction facilitated by running with TRIM disabled.

I have some interesting facts about this TRIM test to share in an upcoming article, so I'll be deliberately vague. For now, understand that there is a reason why drives do better with or without TRIM, and the answer isn't what you'd expect.

Instantaneous

But I also want results for the instantaneous average of my TRIM test. How does the drive fare servicing writes with and without TRIM during each 100,000-command window? The solid teal line represents IOPS across the entire trace, without TRIM. The dotted purple line is with TRIM. Each data point represents write IOPS per 100,000-command test reporting period. Note the difference in the run with TRIM and without.

I didn't want to bowl you over with complicated charts, so I'll focus on the 128 GB 850 Pro. As we saw in the bar chart, performance increases with capacity, and as you ascend the 850 Pro family, the difference between running with TRIM on and off shrinks. This consequently becomes the most illustrative example of the command in play.

The plucky 128 GB 850 Pro does well with and without TRIM. But as the test progresses, the command is needed to service those big spikes towards the end.

Throughput

We collect and report the total throughput of each drive in the NCQ with TRIM test. It's one number that helps capture overall performance in the test.

The 850 Pros don't land in the very front. Then again, look at the drives they're going up against. This is Intel's wheelhouse, and the same technology featured in its enterprise-class SSDs shows up in that compelling SSD 730, too.

9. Testing The DevSlp Power State With Some New Gear

Sometimes I find it unfortunate that most of our storage analysis is in the context of desktop PCs, where the power consumption of an SSD doesn't really matter. The topic is far more meaningful in the enterprise and mobile spaces though, so I find it critically important to benchmark power thoroughly and as precisely as possible.

On a laptop, every milliwatt matters. So much so, in fact, that Intel's Haswell-based CPUs and corresponding chipsets on the mobile side support a new mode for reducing SSD power consumption. DevSlp, or device sleep, is a sideband signal sent to the storage device to indicate that it should drop into a super-low power state. Essentially, everything that can be off is.

This is a great way to get a little extra battery life out of an Ultrabook (particularly in light of Intel's targets for runtime and standby connectivity). But you do pay a price: it takes longer to enter and exit the DevSlp state. Granted, the delay is less than powering the SSD down and back up as needed, a process that can take seconds. Worse, an drive may use substantial amounts of power as it's readied again. DevSlp should need only 50 ms in contrast, along with a few milliwatts.

To measure power consumption in a DevSlp state, we need two things. First is an Ultrabook with a Haswell-based CPU on a compatible platform. I'm using Lenovo's ThinkPad T440s. It's reasonably versatile, including a 2.5" SATA bay and two M.2 slots (for M.2 2242s) wired to the PCH's SATA ports. I typically don't need more than one slot, but it's nice to have the option at least.

The second item is a test platform able to initiate the DevSlp command, measure the current draw, and record the results. To do that, ULINK Technologies sent over some hardware designed expressly for this purpose. I've been using the company's DriveMaster software and SATA/SAS power hubs for a year now, and they confer a spectacular amount of control over what drives under test do. In this case, DevSlp testing is made possible in a way that's informative and easy to manage.

Using a test script to record amperage and issue the appropriate commands, this is what we end up with:

This is an example from my Plextor M6S review. The test script begins at active idle, then issues write commands (the first big increase in power). After 20,000 I/Os, the drive gets issued the DevSlp signal (denoted by the vertical purple bars). In this DevSlp zone, it takes a few tens of milliseconds before the drive enters DevSlp as commanded, but it stays at the state using just 2.5 mW until DevSlp exits (noted by the second purple bar). More I/O is then issued, and then it's back to idle before the script ends. The results are recorded in milliamps, and I convert to watts.

Power Testing With DevSlp

I'll keep it short and sweet this time around, mostly since the next page is going to include all of the power testing. For now, here are some DevSlp and slumber figures for a handful of drives in the lab.

The ULINK DevSlp test platform's instrumentation is sensitive enough to measure thousandths of an amp. Multiply the 2.5" SSDs by an estimate of the PSU's 5 V rail and we get mW from mA. For mSATA- and M.2 SATA-based SSDs, we multiply by the 3.3 V rail's estimated current value. 

In order to get into DevSlp, my DriveMaster 2012 script first enables the device sleep feature bit. From there, it steps through power management modes, finally entering DevSlp from slumber. This has a specific name: DESO. It stands for DevSlp Entry from Slumber Only. So, the last stop before DevSlp is slumber.

DevSlp Measurements

First up, here's the wattage these drives consume in DevSlp. Actually, make that mW; there are way too many zeros involved if I try to report the figures in watts. Remember that when you compare these numbers to the measurements on the next page.

Samsung's spec claims the 850 Pro drops to 2 mW in DevSlp. But that's for the 256 GB model, I suspect. Our tools measure just 1 mW (.2 mA) on the 128 GB version and 2.5 mW on the 1 TB drive.

Compare that data to the 1 TB 840 EVO in mSATA trim (the 2.5" 840 EVOs don't support DevSlp). With the same controller, I get 15.3 mW (5.1 mA * ~3.3 V). Clearly, DevSlp is something vendors are starting to optimize for now. The other handful of drives we have on-hand with support range from 2.5 to 5 mW. I didn't even know the M500 supported DevSlp, but it does, and it works.

Lowest Slumber

This is arguably a more relevant measurement, since it affects a greater number of users. Lowest slumber state power consumption matters to laptops not using connected standby. If DIPM is enabled, these numbers are what you'd see once the host system or device decides to snooze. The same behavior isn't desirable in a desktop setting, which is why we previously only showed active idle numbers.

Not every drive supports every power management trick in the book, so this is the lowest measured non-DevSlp power state.

The Samsungs don't use much power here, either. The two 1 TB MEX-equipped models demonstrate just 38 mW of consumption in slumber, just over twice what the 840 EVO mSATA uses in DevSlp. But that's 38x as much as the 128 GB 850 Pro needs in its deepest sleep state.

Considering that maximum power for the 1 TB drive is around 3000 mW, this is hot stuff. Power management is for real.

10. Results: Power Testing

Active Idle Power Consumption

Idle consumption is the most important power metric for consumer and client SSDs. After all, solid-state drives complete host commands quickly and then drop back down to idle. Aside from the occasional background garbage collection, a modern SSD spends most of its life doing very little. Enterprise-oriented drives are more frequently used at full tilt, making their idle power numbers less relevant. But this just isn't the case on the desktop, where the demands of client and consumer computing leave most SSDs sitting on their hands for long stretches of time.

Active idle power numbers are critical, especially when it comes to their impact on mobile platforms. Idle means different things on different systems, though. Pretty much every drive we're testing is capable of one or more low-power states, up to and including the DevSlp stuff we covered on the previous page.

Idle power consumption is stated as up to .4 W maximum by Samsung, but I see a steady .28 W each for the 128, 256, and 1024 GB 850 Pro. Rounding gives the 256 GB model an extra hundredth.

PCMark 7 Average Power Consumption

If we log power consumption through a workload, even a relatively heavy one, we see that average use is still pretty close to the idle numbers. Maximum power may spike fiercely, but the draw during a PCMark 7 run is light. You can see the drives fall back down to the idle "floor" between peaks of varying intensity.

All three capacities peak between 2 and 3 W, and each drops down to .3 W in between jobs. Because each drive comes equipped with a different amount of flash, they don't line up exactly. But it's easy to see that the 850 Pro is easy on power, even during heavy use.

The maximum power consumption of Samsung's 1 TB 850 Pro is around 3 W. Exact figures for the other capacities are given. The transition is smooth, though. I believe that different die sizes are being used for the smaller drives. But without knowing more about V-NAND, only Samsung's tight-lipped engineers know for sure.

11. Results: Latency And Performance Consistency

Normally, I'd just use a garden variety workload generator for testing latency and performance consistency. It's simple enough that you could train a medium-smart kitten to do this stuff. But I like ULINK's platform; it gives me a common foundation for running benchmarks. That's not a big deal for some metrics, but it's more important for others. I have to use several different machines for testing with DriveMaster, due mostly to the length of time these workloads eat up. Two comparable systems deliver identical performance with DM2012.

Latency

ULINK's Latency test does things in a slightly different way than I've presented in the past. First, the drive is secure erased. Next, it is filled twice sequentially, then once with random writes. Immediately after, the latency test begins.

This script tests sequential one-sector accesses, sequentially testing 512 bytes. Do that 2 million times for reads and two million times more for writes, and presto. You get one result for reads and another for writes in the blink of an eye.

Samsung's hardware looks impressive. Not only do the 850 Pros serve up low latencies, but the 840 EVO and 840 Pro also hold their own. It takes the other contenders much longer to complete reads and writes under the same conditions, ending with Crucial's MX100. We'll do more of this testing on the drives waiting for reviews, but you can already see that going the budget route has negative implications.

Performance Consistency

This is another test I'd typically run differently. I don't even really like the phrase performance consistency. But I do like ULINK's PerfCon test. It offers six test cases: 4 KB read, 4 KB write, and 50% read/50% write in both aligned and unaligned boundaries. I won't bore you with the outcomes of all six test cases. Instead, I'll stick to the 50% read/50% write 4 KB aligned measurement.

The trick here is that after the same preconditioning used above, 1000 data points are generated one second apart. The 95.5th-percentile slowest result is divided by the average IOPS to yield one performance number reflecting consistency as a percentage. Absolute transactional performance is factored out of that number, so I display average, minimum, and maximum IOPS in a separate chart.

As you can see, the average, maximum, and minimum IOPS results are more even towards the top. In this 4 KB random workload split between reads and writes, steady state 4 KB write performance gets bolstered by reads. But it's more difficult for some drives to maintain when the workload is mixed.

Sure enough, Intel's SSD 730 ends on top, owing to its high average I/O performance. The Samsungs filter in behind. There's almost 2.5x as much difference between the slowest  and fastest averages. But how do the percentages play out?

The Vector exhibits some strange results, but 65% isn't bad for the MX100. Intel's SSD 730 attains an impressive 87%, and the 850 Pro contingent (led by the 256 GB capacity point) score a win. Each is above 91%, which is just great. Even without additional over-provisioning, the 850 Pros dominate in our look at consistency.

12. SATA Is Maxed, But The 850 Pro Still Pushes Faster

The all-encompassing brilliance of Samsung's 850 Pro is muted somewhat by the fact that SATA 6Gb/s is restrictive. The 850 Pro works within the interface's constraints, though. It's well-rounded and fast, no doubt. But most of the SSDs we've reviewed lately land within a stone's throw of this drive, so it's hard to declare it the undisputed F.O.A.T (fastest of all time, to bastardize an LL Cool J album title). Really, the data speaks for itself though, and in a majority of tests, Samsung lands on top.

We'll continue lamenting the damper a 6 Gb/s interface puts on new SSD launches until alternative connections become more common. In the meantime, 3D V-NAND does, in fact, appear to benefit performance and power measurements alike.

Speaking of, our readings in DevSlp are spectacular, though it's hard to get amped up about insanely low power readings, particularly on the desktop where sleep states add unwanted latency. As the feature becomes more prolific in the mobile space, I'd expect Samsung's 850 Pro to become a favorite choice in notebooks, though.

Similarly, most enthusiasts won't avail themselves of the Pro's encryption capabilities. I'm glad they're supported in hardware, but I suspect a lot of 850 Pros will end up in multi-drive arrays where encryption would be enabled through a hardware RAID controller. Otherwise, it's not clear just how much of the power user space burns for the sweet, obfuscated fruit of full disk encryption, Microsoft's eDrive standard, or Opal 2.0.

The 850 Pro's centerpiece is its 3D V-NAND, which is said to make a monumental contribution to endurance. That's more difficult to test (and come away with useful information). Yes, performance is better because of the technology, and we benchmarked that. Power consumption is also better, and I measured that using some pretty high-end equipment. Unfortunately, TBW (terabytes written) specifications, as confusing as they are, tend to involve other information we simply don't get from Samsung. Intel's SSD 730 can withstand up to 70 GB per day through its warranty term, while the 850 Pro should land around 40 GB daily. But that's running the math with a 10-year warranty in mind. Recalculate for a five-year term and the 850 Pro gets a healthier 80 GB/day endurance spec.

Of course, the method Samsung uses to calculate its figures isn't divulged. I just don't think it matters, though. Once upon a time, I put almost seven million gigabytes on a 256 GB Samsung 830. All that means is your approach to testing endurance matters as much as how you interpret the results. Think of it like this, though: in the twentieth century, the average life expectancy for Americans rose. Seemingly minor illnesses were treated more effectively as the century progressed, so the number of deaths from cancer and heart disease rose. Dumb stuff stopped killing us as often, leaving more of the population alive long enough to develop a more terminal condition. We're all going to go sometime, from something. And in the same way, all SSDs die on a long enough timeline. If it's not from something preventable like poor NAND management, sketchy component choice, or flaky firmware, the drive may last long enough for endurance to become its undoing. For a majority of us, though, that's not something worth losing sleep over.

Do keep in mind that a decade of warranty coverage sounds awesome, but it's limited to the TBW figure. You get 10 years or 150 TBW, whichever happens first. Increasing coverage by five years probably won't cost Samsung much in the long run based on that write specification. Also, the TBW rating for each drive is the same. That means the lowest common denominator (the tiny 128 GB model) was likely the guinea pig for Samsung's calculations. I have a hard time believing the 1 TB 850 Pro wouldn't outstrip the cited number. 

It all comes together in an impressive package. Squeezing every last bit of headroom from SATA 6Gb/s may seem like a fool's errand, but it's that last percentage point that puts these drives in a pole position. The spread between an average SSD and the 850 Pro isn't enormous, and even less so if you're looking at competing high-end offerings. Many storage tests tend to exaggerate performance deltas, after all. If I swapped your Radeon R9 290X for a 290, you might not even notice. Similarly, if I pulled your existing 840 EVO and upgraded you to an 850 Pro, the difference would likely be imperceptible. It's only when we really push these devices to their limit that certain SSDs shine. Like the SSD 730 from Intel, Samsung's new 850 Pro holds its own in those taxing situations.

Frankly, most readers (even the enthusiasts) won't need the 850 Pro, particularly given relatively steep pricing. Most of us with desktops and notebooks are well-served by superb offerings battling it out at the budget end of the market. There are too many options going for less than $.50/GB, including certain configurations of the 840 EVO, to jump all over flagships selling for twice as much. But then there are the most hardcore users who willingly pay handsomely for the fastest CPUs and graphics cards. They run entry-level servers, edit high-resolution video, build performance-sensitive RAID arrays, and so on. They're the ones who'll find what the 850 Pro can do most interesting.

I can say that Samsung turns the dial as high as it'll go for SATA 6Gb/s. There's not much room left to innovate until we start seeing versions of the 850 designed for alternative interfaces. But somehow, the 850 Pro on my bench right now pushes a little more performance across the board. It's a triumph in its own way. Only power users need apply, though (at least until prices drop a bit). Other vendors want the distinction Samsung claims for itself today. For the moment, however, this 850 Pro gets to keep the crown warm.