Samsung’s 850 EVO solid-state drive reviewed

Samsung’s 850 EVO SSD has been a long time coming. We got our first hint of its existence in July, when the firm revealed that it was developing a TLC version of its 3D V-NAND. More details on the three-bit variant were released in August, when Samsung confirmed that the flash was destined for the 850 EVO. Then, in October, mass production of TLC V-NAND chips began, setting the stage for today’s formal unveiling.

The new EVO has some Shaq-sized shoes to fill. Samsung claims that its predecessor, the 840 EVO, is the best-selling SSD on the market. We haven’t seen sales figures to support that assertion, but the EVO does have more than double the number of Amazon user reviews of its peers. An overwhelming majority of those—4,540 as I write this—give the drive a five-star rating.

Our expectations are especially high because the 850 EVO’s elder sibling, the 850 Pro, is the fastest SATA SSD we’ve ever tested. Samsung’s latest hotness is based on the same core technology, just with an extra bit per cell and some clever caching to pick up the slack. And it’s backed by a higher endurance rating and longer warranty than most SSDs.

In the realm of SATA SSDs, the 850 EVO is kind of a big deal. Let’s see what it can do.

TLC in three dimensions
The defining element of the 850 EVO is easily its three-dimensional flash. This new, three-bit implementation has the same 32-layer structure as the MLC V-NAND in the 850 Pro. The cell geometry is the same, and the die area should be similar, but the chips are fabbed at a different facility. Samsung doesn’t mass-produce V-NAND and then sort the results into MLC and TLC bins; the chips in the 850 EVO were born to host three bits per cell

Indeed, this entire generation of V-NAND seems tailor-made for TLC. The individual chips offer 128Gb of capacity with three bits per cell, a much rounder number, at least in PC storage terms, than their 86Gb MLC payload. That makes perfect sense, because V-NAND’s strengths nicely offset TLC’s weaknesses.

Three-bit flash has lower write performance and endurance than its two-bit counterpart. The extra bit is at fault: it requires the cell to differentiate between twice as many possible values within the same limited voltage range. Writing and verifying data is more difficult—and slower—as a result. TLC’s higher bit density makes the cells more sensitive to normal flash wear and more prone to interference from adjacent cells.

V-NAND tackles those weaknesses in several ways. First, it trades the floating gates typically found in planar NAND for a 3D charge-trap that’s inherently more robust. This structure stores electrons in a trapping oxide that wraps around a vertical electron channel. The oxide is an insulator, so any physical damage only affects electrons in the immediate vicinity. With floating gates, a breach in the oxide can drain the entire contents of the cell, rendering it useless.

Stacking cells vertically also helps V-NAND’s endurance. Instead of increasing the bit density by making the cells smaller, and thus weaker, Samsung lays down multiple layers based on a fairly large 40-nm planar geometry. The company claims this coarser 2D process puts enough distance between the cells to completely eliminate interference along the horizontal bit lines. Samsung says there’s enough space between the layers to virtually eliminate interference along the vertical word lines.

With less interference to worry about, V-NAND can employ a simpler programming algorithm that improves write performance. It can also exploit the inherent robustness of the trapping oxide to program the cells more aggressively—or write less aggressively to extend the life of the flash.

TLC V-NAND still has an extra bit per cell, so it can’t match the peak speed or endurance of its MLC counterpart. However, Samsung claims the EVO’s flash is good enough to keep up with the planar MLC NAND found in most SSDs.

The endurance ratings certainly put the EVO in MLC territory. According to the official specs, the 120GB and 250GB versions are good for 75TB of total writes, while the 500GB and 1TB flavors can withstand 150TB. Impressively, the larger EVOs have the same endurance rating as the 850 Pro.

We don’t have any V-NAND units in our ongoing endurance test, but that experiment has taught us that SSDs can last much longer than their specifications promise. Common sense also tells us that point is largely academic. PC users typically write no more than a few terabytes a year, which is much less than modern drives are rated to survive.

Note that the specs table lists two write speeds for each drive capacity. The first figure describes the performance of TurboWrite, a flash-based caching scheme inherited from the 840 EVO. TurboWrite addresses a portion of the flash in a single-bit SLC configuration that offers higher performance and endurance than even MLC setups. Incoming writes are funneled into this cache and then moved to TLC storage during idle periods. If the cache is filled before it can be flushed, inbound data proceed directly to three-bit cells, and write performance drops accordingly. The second set of write speed figures in the table above refer to TLC writes.

We’re waiting on Samsung to confirm the cache sizes for the various capacities, but if the implementation is the same as in the 840 EVO, the 1TB drive has a 9GB cache, the 500GB has 6GB, and the others have 3GB. That’s not loads of storage, but it should be sufficient to cover the shorter, bursty writes typical of client workloads. The 850 EVO’s performance specs also suggest caching effectively offsets the performance penalty associated with lower-capacity SSDs. Despite lacking sufficient flash dies to exploit the controller’s highly parallel NAND interface, the 120GB and 250GB drives nearly match the TurboWrite peaks of the higher-capacity EVOs.

In a surprise twist, the 850 EVO 1TB pictured above uses the same triple-core MEX controller as the 850 Pro, while the rest of the family taps a newer MGX chip with only two cores. Samsung contends that V-NAND’s raw performance is good enough for the smaller drives to reach top speed with one fewer core. The greater “hardware automation” introduced with the MEX generation likely lessens the need for additional general-purpose horsepower, as well.

Dropping a core saves cost and power, both important factors in the solid-state sphere. SSD makers are increasingly competing on price, and Samsung is evidently shaving pennies wherever it can. Just look at the tiny circuit board used for the 250GB variant:

For additional perspective, here’s how the two drives look inside their 2.5″ shells:

Samsung packs a lot of other goodness into these tiny packages. The 850 EVO’s 256-bit AES encryption is compliant with the TCG Opal 2.0 and IEEE 1667 standards, meeting the requirements for Microsoft’s eDrive system. The drive supports the ultra-low-power DevSleep state, and it can throttle its clocks to prevent overheating in toasty environments. Then there’s the five-year warranty, a premium perk that defies the series’ mid-range pricing.

The EVO works with Samsung’s excellent SSD Magician utility, which monitors health stats, performs firmware updates, and migrates old data to new drives. That utility also includes RAPID mode, an optional, DRAM-based caching scheme similar to Windows’ SuperFetch mechanism.

We haven’t tested DRAM caching in this review (there’s a full suite of RAPID results in our 850 Pro coverage), but we have put the 850 EVO 250GB and 1TB through their paces against a deep field of competitors. Our performance analysis begins on the next page.

CrystalDiskMark — transfer rates
First, we’ll tackle sequential performance with CrystalDiskMark. This test runs on partitioned drives with the benchmark’s default 1GB transfer size and randomized data.

We’ve color-coded the results to make the 850 EVOs easier to spot. We’ve also highlighted the other Samsung SSDs, coloring the TLC and MLC drives in different shades.

Both 850 EVOs are near the front of the pack in our CrystalDiskMark tests. Even the 250GB unit delivers strong write performance, which is really no surprise. The 1GB test should fit easily inside the TurboWrite buffer.

HD Tune — random access times
Next, we’ll turn our attention to random access times. We used HD Tune to measure access times across multiple transfer sizes. SSDs have near-instantaneous seek times, so it’s hard to graph the results on the same scale as mechanical drives. The WD Black and Seagate SSHD will sit out this round to focus our attention on the SSDs.

The 1MB random write test trips up some of the SSDs, but not the 850 EVOs, which are competitive throughout. To be fair, virtually all the SSDs have comparable access times in the other tests. Tiny fractions of a millisecond separate most of the drives in most of these tests.

TR FileBench — Real-world copy speeds
FileBench, which was concocted by TR’s resident developer Bruno “morphine” Ferreira, runs through a series of file copy operations using Windows 7’s xcopy command. Using xcopy produces nearly identical copy speeds to dragging and dropping files using the Windows GUI, so our results should be representative of typical real-world performance. We tested using the following five file sets—note the differences in average file sizes and their compressibility. We evaluated the compressibility of each file set by comparing its size before and after being run through 7-Zip’s “ultra” compression scheme.

The names of most of the file sets are self-explanatory. The Mozilla set is made up of all the files necessary to compile the browser, while the TR set includes years worth of the images, HTML files, and spreadsheets behind my reviews. Those two sets contain much larger numbers of smaller files than the other three. They’re also the most amenable to compression.

To get a sense of how aggressively each SSD reclaims flash pages tagged by the TRIM command, the SSDs are tested in a simulated used state after crunching IOMeter’s workstation access pattern for 30 minutes. The drives are also tested in a factory fresh state, right after a secure erase, to see if there is any discrepancy between the two states. There wasn’t much of one with the 850 EVOs, so we’re only presenting the used-state scores.

The 850 EVO performs very well versus its peers, posting competitive copy speeds with both larger and smaller files. Despite having fewer cores in its controller and fewer dies at its disposal, the 250GB model holds its own in FileBench. In some of the tests, it’s even faster than its 1TB counterpart and the 850 Pro.

TR DriveBench 2.0 — Disk-intensive multitasking
DriveBench 2.0 is a trace-based test comprised of nearly two weeks of typical desktop activity peppered with intense multitasking loads. More details on are available on this page of our last major SSD round-up.

We measure DriveBench performance by analyzing service times—the amount of time it takes drives to complete I/O requests. Those results are split into reads and writes.

Impressively, the 850 EVO 1TB nearly matches the mean service times of the 850 Pro. The 250GB drive isn’t as quick, but it still represents a substantial step up from the equivalent 840 EVO. The differences between the two generations are particularly acute with writes.

All the SSDs execute the vast majority of DriveBench requests in one millisecond or less—too little time for end users to perceive. We can also sort out the number of service times longer than 100 milliseconds, which is far more interesting data. These extremely long service times make up only a fraction of the overall total, but they’re much more likely to be noticeable.

The 850 EVOs slog through fewer extremely long reads than their planar peers. They barely log enough longer writes to show up on the radar, which is especially impressive for the 250GB version. The 840 EVO 250GB suffers more than a thousand writes over 100 milliseconds—150 times more than its replacement.

IOMeter
Our IOMeter workload features a ramping number of concurrent I/O requests. Most desktop systems will only have a few requests in flight at any given time (87% of DriveBench 2.0 requests have a queue depth of four or less). We’ve extended our scaling up to 32 concurrent requests to reach the depth of the Native Command Queuing pipeline associated with the Serial ATA specification. Ramping up the number of requests also gives us a sense of how the drives might perform in more demanding enterprise environments.

We run our IOMeter test using the fully randomized data pattern, which presents a particular challenge for SandForce’s write compression scheme. We’d rather measure SSD performance in this worst-case scenario than using easily compressible data.

There’s too much data to show clearly on a single graph, so we’ve split the results. You can compare the 850 EVO’s performance to that of the competition by clicking the buttons below each graph.

Instead of presenting the results of multiple access patterns, we’re concentrating on IOMeter’s database test. This access pattern has a mix of read and write requests, and it’s similar to the file server and workstation tests. The results for these three access patterns are usually pretty similar. We also run IOMeter’s web server access pattern as part of our standard suite of tests, but it’s made up exclusively of read requests, so the results aren’t as applicable to real-world scenarios. Our own web servers log a fair amount of writes, for example.

Although the 850 EVOs don’t match the I/O throughput of the 850 Pro, they do deliver higher transaction rates than their ancestors. The 250GB and 1TB incarnations are evenly matched up to four concurrent requests, after which the former starts to plateau. The smaller drive is still competitive with most SSDs, but its larger sibling scales higher, up to a queue depth of 16.

Regardless of the queue depth, the 850 EVOs are well behind the standouts in this test. The Intel 730 Series is an absolute monster across the board, and some of the OCZ drives offer even higher transaction rates under the heaviest load.

Boot duration
Before timing a couple of real-world applications, we first have to load the OS. We can measure how long that takes by checking the Windows 7 boot duration using the operating system’s performance-monitoring tools. This is actually the first test in which we’re booting Windows off each drive; up until this point, our testing has been hosted by an OS housed on a separate system drive.

Level load times
Modern games lack built-in timing tests to measure level loads, so we busted out a stopwatch with a couple of titles.

To no one’s surprise, the 850 EVO’s load times aren’t much different from those of the other SSDs. Most of the drives are separated by less than a second.

As soon as this stream of new releases subsides, we’ll be switching to an all new collection of load-time tests. We have faster test rigs and PCIe SSDs that may provide more interesting results.

Power consumption
We tested power consumption under load with IOMeter’s workstation access pattern chewing through 32 concurrent I/O requests. Idle power consumption was probed one minute after processing Windows 7’s idle tasks on an empty desktop.

The 850 EVO 250GB has particularly low power consumption at idle and under load. Its terabyte twin requires more juice, but that drive still has lower idle power than anything in its weight class. All the other terabyte SSDs draw less power under load, though.

That’s it for performance. If you’re curious about the other SSDs in this review or about how we conduct our testing, hit up the methods section on the next page. Otherwise, feel free to skip ahead to the conclusion.

Test notes and methods
Here’s a full rundown of the SSDs we tested, along with their essential characteristics.

	Flash controller	NAND
Adata Premier SP610 512GB	Silicon Motion SM2246EN	20-nm Micron sync MLC
Adata Premier Pro SP920 512GB	Marvell 88SS9189	20-nm Micron sync MLC
Corsair Force GT 240GB	SandForce SF-2281	25-nm Intel sync MLC
Corsair Force LS 240GB	Phison PS3108-S8	19-nm Toshiba Toggle MLC
Corsair Neutron 240GB	LAMD LM87800	25-nm Micron sync MLC
Corsair Neutron GTX 240GB	LAMD LM87800	26-nm Toshiba Toggle MLC
Corsair Neutron XT 240GB	Phison PS3110-S10	A19-nm Toshiba Toggle MLC
Crucial M500 240GB	Marvell 88SS9187	20-nm Micron sync MLC
Crucial M500 480GB	Marvell 88SS9187	20-nm Micron sync MLC
Crucial M500 960GB	Marvell 88SS9187	20-nm Micron sync MLC
Crucial M550 256GB	Marvell 88SS9189	20-nm Micron sync MLC
Crucial M550 512GB	Marvell 88SS9189	20-nm Micron sync MLC
Crucial M550 1TB	Marvell 88SS9189	20-nm Micron sync MLC
Crucial MX100 256GB	Marvell 88SS9189	16-nm Micron sync MLC
Crucial MX100 512GB	Marvell 88SS9189	16-nm Micron sync MLC
Intel 335 Series 240GB	SandForce SF-2281	20-nm Intel sync MLC
Intel 520 Series 240GB	SandForce SF-2281	25-nm Intel sync MLC
Intel 730 Series 480GB	Intel PC29AS21CA0	20-nm Intel sync MLC
OCZ Vertex 4 256GB	Indilinx Everest 2	25-nm Micron sync MLC
OCZ Vertex 450 256GB	Indilinx Barefoot 3 M10	20-nm Micron sync MLC
OCZ Vertex 460 240GB	Indilinx Barefoot 3 M10	19-nm Toshiba Toggle MLC
OCZ ARC 240GB	Indilinx Barefoot 3 M10	A19-nm Toshiba Toggle MLC
SanDisk Extreme II 240GB	Marvell 88SS9187	19-nm SanDisk Toggle SLC/MLC
Samsung 840 Series 250GB	Samsung MDX	21-nm Samsung Toggle TLC
Samsung 840 EVO 250GB	Samsung MEX	19-nm Samsung Toggle TLC
Samsung 840 EVO 500GB	Samsung MEX	19-nm Samsung Toggle TLC
Samsung 840 EVO 1TB	Samsung MEX	19-nm Samsung Toggle TLC
Samsung 840 Pro 256GB	Samsung MDX	21-nm Samsung Toggle MLC
Samsung 850 EVO 250GB	Samsung MGX	32-layer Samsung V-NAND TLC
Samsung 850 EVO 1TB	Samsung MEX	32-layer Samsung V-NAND TLC
Samsung 850 Pro 512GB	Samsung MEX	32-layer Samsung V-NAND MLC
Seagate 600 SSD 240GB	LAMD LM87800	19-nm Toshiba Toggle MLC
Seagate Desktop SSHD 2TB	NA	24-nm Toshiba Toggle SLC/MLC
WD Caviar Black 1TB	NA	NA

The solid-state crowd is augmented by a couple of mechanical drives. WD’s Caviar Black 1TB represents the old-school hard drive camp. Seagate’s Desktop SSHD 2TB is along for the ride, as well. The SSHD combines mechanical platters with 8GB of flash cache, but like the Caviar Black, it’s really not a direct competitor to the SSDs. The mechanical and hybrid drives are meant to provide additional context for our SSD results.

Since you’ve made it this far, here are a couple more shots of the drives we tested:

We used the following system configuration for testing:

Thanks to Asus for providing the systems’ motherboards and graphics cards, Intel for the CPUs, Corsair for the memory and PSUs, Thermaltake for the CPU coolers, and Western Digital for the Caviar Black 1TB system drives.

We used the following versions of our test applications:

• Intel IOMeter 1.1.0 RC1
• HD Tune 4.61
• TR DriveBench 1.0
• TR DriveBench 2.0
• TR FileBench 0.2
• Qt SDK 2010.05
• MinGW GCC 4.4.0
• Duke Nukem Forever
• Portal 2

Some further notes on our test methods:

• To ensure consistent and repeatable results, the SSDs were secure-erased before almost every component of our test suite. Some of our tests then put the SSDs into a used state before the workload begins, which better exposes each drive’s long-term performance characteristics. In other tests, like DriveBench and FileBench, we induce a used state before testing. In all cases, the SSDs were in the same state before each test, ensuring an even playing field. The performance of mechanical hard drives is much more consistent between factory fresh and used states, so we skipped wiping the HDDs before each test—mechanical drives take forever to secure erase.

• We run all our tests at least three times and report the median of the results. We’ve found IOMeter performance can fall off with SSDs after the first couple of runs, so we use five runs for solid-state drives and throw out the first two.

• Steps have been taken to ensure that Sandy Bridge’s power-saving features don’t taint any of our results. All of the CPU’s low-power states have been disabled, effectively pegging the 2500K at 3.3GHz. Transitioning in and out of different power states can affect the performance of storage benchmarks, especially when dealing with short burst transfers.

The test systems’ Windows desktop was set at 1280×1024 in 32-bit color at a 75Hz screen refresh rate. Most of the tests and methods we employed are publicly available and reproducible. If you have questions about our methods, hit our forums to talk with us about them.