Samsung SM951 128GB And 256GB SSD Review

Samsung's SM951 is the hottest storage product on the market. Today we look at the two smaller versions, both of which still perform better than SATA.

Introduction

The SM951 story has developed since we first shucked a drive from a Lenovo X1 Carbon Gen 3 two months ago. Resellers are selling them at better-than-expected prices. Samsung also announced a new NVMe model appropriately called the SM951-NVMe. The company's 512GB SM951 offers attractive performance for its price, but many waiting for an even better value want information on the smaller capacities before jumping into the world of PCIe-based SSD performance.

If you read our first SM951 review, then you already know the drive we pulled from Lenovo's Ultrabook failed to reach Samsung's SM951 performance specification for sequential reads. At the time, we suspected that Lenovo was limiting the drive to conserve battery power. With more samples in our hands, we now know the issue is Lenovo-specific, and it doesn't affect drives sold by other vendors.

Today we're testing the 128 and 256GB versions of Samsung's SM951 next to the XP941, which officially surfaced in 2014. Back then, the XP941 was the only native PCIe SSD on the market. Marvell has since joined the fight with two native PCIe controllers that will also be represented in the benchmark charts today.

Technical Specifications

Loading...

The three SM951 drives we're testing use AHCI, and are not the rare NVMe interface. In the coming weeks, NVMe-based models will begin surfacing with channel availability soon after. Several online shops already list pre-order status for the NVMe versions, though we've yet to hear of anyone with product on-hand.

The 512GB SM951 we have is Lenovo-specific, with the part number MZHPV512HDGL-000L1. That L1 at the end is a reference to Lenovo. The last five digits indicate the region the drive is sold in. Common codes there are 00001 to 00004, with HP and Lenovo assigned special modifiers. We can also assume that Dell has its own code, since the company plans to release Ultrabook models with SM951 SSDs. 

We already know that Lenovo's SM951 has limited sequential read performance. In our testing, we achieved roughly 1700 MB/s, while the 256GB SM951 we received from RamCity is capable of the full 2150 MB/s. The RamCity 128GB sample achieved 2050 MB/s. Because of this, we've noted the Lenovo model in our charts as a Lenovo SM951 512GB. Soon, we'll have a fresh 512GB drive to test for an upcoming editorial that covers the best PCIe and SATA 6Gb/s SSDs, as well as the fastest hard drives on the market today.

The 512GB and 256GB SM951 models share nearly identical performance specifications. The 256GB model loses some write performance. And the 128GB version drops to "just" 600 MB/s sequential reads. At least it retains the same 70,000 random read IOPS rating as the other models.

Pricing And Accessories

The SM951s come from Samsung's SSI group, which sells to OEM customers. These products are normally not available to end-users. But demand for the company's SSDs compelled Samsung to let a limited supply bleed through to a few resellers large enough to have accounts with key distributors. Vendors spec their own warranty terms, conditions, pricing and even packaging.

We get all of our SSI products from RamCity, an Australian-based company. RamCity sells direct and also has an Amazon account where products are stored and shipped from North America. The SM951s are hot items, so prices are subject to change based on availability. The Australian site advertises in Australian Dollars with Australian VAT taxes that do not apply to orders coming from other countries.

A Closer Look At The SM951 SSDs

Both of the drives we're testing today share the same layout, controller and DRAM. The main difference between them is that the number of dies per package doubles on the 256GB SM951.

All Samsung SM951 SSDs ship in the M.2 2280 form factor. The firmware installed on both of our samples is BXW25000. The Lenovo model's firmware is BXW22L00.

The two lower-capacity SM951s use just two NAND flash packages.

The smaller drives are also single-sided; all of their surface-mount components are on one side. The 512GB model is double-sided. This can be important, since some companies use a single-sided connector that sits closer to the motherboard.

All three capacities employ the same Samsung UBX controller. 

The two lower-capacity models use the same DRAM, while the 512GB model ships with twice as much DRAM to cache page table data. 

The flash is the same 1xnm across all three models. Only the number of dies per package increases alongside capacity. The two drives we're testing today sport two packages. We suspect Samsung uses quad-plane flash, and that is how it manages to reach this level of performance.

MORE: Best SSDs For The Money
MORE: All Storage Articles

MORE: Latest Storage News

MORE: Storage in the Forums

Sequential Read

To read about our storage tests in-depth, please check out How We Test HDDs And SSDs. Four-corner testing is covered on page six.

Right from the start, both of the SM951s from RamCity outperform the SM951 from Lenovo. Normally we'd expect the higher-capacity SSD to perform better thanks to its increased parallelism. But in this case, the difference between the 256GB and 512GB models is large enough to fit a SATA 6Gb/s SSD in. You could park the peak performance of an 850 Pro in the gap!

It appears that Lenovo limits the SM951 for its Ultrabooks. This would reduce power consumption and the amount of heat generated inside its compact chassis.

Sequential Write

The difference in sequential write performance between the three SM951s is significant. That wave in the 512GB model's output could be due to thermal throttling. We ran the test several times and came away with the same pattern. The SM951's controller runs cooler than the XP941's. As a major manufacturer, Lenovo could have asked Samsung to build a drive with a lower TDP for its Ultrabooks.

Random Read

Random performance ties in with latency, what makes your computer feel fast or not. It's hard to believe, but the latest M.2-attached SSDs exceed 12,000 random read IOPS at a queue depth of one. On the line chart, we see that all three SM951s outperform Samsung's specifications. And again, the two lower-capacity models deliver better performance than Lenovo's.

Random Write

At low queue depths, all three SM951 drives are in a league of their own. Removing the Lenovo SM951 from the group, the 256GB implementation delivers nearly 10,000 more random write IOPS at a queue depth of one. Samsung is one of the few SSD manufacturers that focuses on low queue depth performance. Samsung is also one of the only companies that tells you what random read and write performance at a queue depth of one to expect.

Mixed Workloads

Our mixed workload testing is described in detail here, and our steady state tests are described here.

Mixed workload performance in client workloads has always been an important consideration. But until recently, it was largely ignored since SATA-based drives can only transmit or receive data at one time. PCIe is bi-directional, but the flash controller also needs to take advantage of two-way communication.

Samsung's XP941 delivers around the same mixed workload performance as the best SATA products. When I first tested it, I introduced our mixed workload testing on client SSDs. Since then, Samsung has worked on improving those results. The SM951 is optimized for sending and receiving data simultaneously, outpacing other M.2-based drives. This is an area where we expect the new SM951-NVMe to really stand out once it becomes available.

128KB Sequential Mixed-Workload Steady State Performance

For this review and for future stories with a workstation twist, we're adding a look at 70%-read steady state mixed sequential transfers.

The line chart shows sequential steady state performance from 100% reads to 100% writes at high queue depths. The two larger SM951s perform well beyond competing products. The 128GB model falls in line with other products on the market when the mix moves closer to a static workload. 

4K Random Write Steady State Performance

As in the previous test, Samsung's 128GB SM951 gets lost in the mix as the higher-capacity models shoot above the crowd. Kingston's HyperX Predator performs well here thanks to Marvell's new Altaplus controller. Still, none of the drives on the chart hold a consistent level like Intel's SSD 750.

PCMark 8 Real-World Software Performance

For details on our real-world software performance testing, please click here.

After watching the SM951 dominate our synthetic tests, we aren't surprised to see it at the top of our real-world benchmarks as well. The results from PCIe-based M.2-attached drives are even tighter than what we gathered from SATA-based drives. This really goes to show that if you are not multitasking or writing enough data to induce steady state conditions, you won't notice a lot of difference between this class of SSDs.

Total Storage Bandwidth

The slight performance differences compound over time. Here we see the converged results, and the SM951s are at the top. But that isn't a surprise. Meanwhile, the XP941s fall back a bit, mixed in with the Marvell-based drives. Samsung appears to be a full generation ahead of Marvell in performance.

PCMark 8 Advanced Workload Performance

To learn how we test advanced workload performance, please click here.

Even though the Lenovo-specific SM951 demonstrates lower read performance than the two lower-capacity models, higher write speed is achieved through parallelization, keeping the 512GB implementation at the top of our chart under heavy and light workloads. The 256GB SM951 falls in line just behind, while the 128GB drive with the fewest number of dies blends in with higher-capacity models using Marvell controllers.

Latency Tests

The two larger SM951s yield the best throughput of the drives on our chart, but the higher-capacity XP941s complete the tasks in less time. The Lenovo-specific SM951 lands just behind last year's OEM superstar from Samsung. Meanwhile, the smaller SM951s take longer to finish the tasks under light workload conditions coming out of heavy workloads.

Lower-capacity Samsung SSDs are hit and miss when it comes to cleaning up the NAND after heavy workloads. Many of the company's products displayed this behavior. What we're seeing is the clean-up process taking longer. New reads and writes happening at the same time as background operations increases latency. I noticed this with the 830- and 840-series drives. The 850s with 3D V-NAND were more resilient, even after especially taxing workloads. The SM951 uses 1xnm 2D planar MLC NAND that evolved from what we found in the 840-series, though.

The light workload set assumes that five minutes between tests is long enough for the drive to tidy up dirty cells. So either it takes longer to clean house, or the house gets dirty faster with the large 128Gb dies. Both the 128GB and 256GB SM951s use just two NAND packages and the 512GB model uses four. This can limit I/O, even with quad-plane packages.

Here we see the 128GB SM951 in a dirty state after a reasonable number of sequential and random writes. At this point in the test process, the drive wasn't being pushed particularly hard. Despite that, 128KB sequential write performance is already down to very low levels. Over time, the internal wear-leveling algorithms will clean the flash and write performance should recover where cells are free.

Notebook Battery Life

For more information on how we test notebook battery life, click here.

After our first SM951 test, Lenovo released a firmware update that lets us adjust the amount of time it takes to drop into slumber. We updated the firmware and set this timer to two minutes. Then, we retested all of the drives for this chart.

Samsung's SM951 is the first client SSD with support for the L1.2 power state defined by PCI-SIG. At full rest, it draws just 2mW. What really surprises us is just how much better Samsung's drives do in this discipline than other SSDs with Marvell controllers.

On battery power, your notebook scales back several system bus clock speeds to save power. In this reduced power state, all of the drives in today's chart perform nearly identically. We sometimes find a product that stands out from the rest, providing better efficiency. For the time being, though, one M.2 PCIe-based SSD fares just as well as the rest.

Final Thoughts

Samsung enjoys a large lead in both controller and flash technology over its competitors. 3D V-NAND stacked 32 layers-high covers the flash. Only Samsung has this. And the UBX isn't Samsung's first PCIe-to-flash controller efficient enough for power-restricted M.2 applications; it's a second-generation part.

Marvell is trying to catch up, but even the second-gen Altaplus controller trails Samsung's effort. SandForce, now part of Seagate, could make some noise at Computex. However, the SF3700 is two years late and we haven't seen anywhere near 2150 MB/s in our behind-the-scenes testing. Phison, JMicron and others plan to release PCIe-based controllers in 2015. We're not holding our breath for Samsung-class performance right out of the gate, though.

Our story today focused on the M.2 form factor, so Intel's SSD 750 wasn't represented. In a few days, an SSD 750 400GB will arrive. Right behind that we should be receiving a trio of SM951-NVMe drives. Once all of the pieces are in place, we'll publish a comparison of high-end NVMe products.

I hate to say that the products in this piece are in a lower-tier, but the SM951-NVMe is going to be more expensive. And we don't expect a large performance increase over the AHCI-based SM951 in client workloads, so the models we tested should provide better value to enthusiasts. Without samples in the lab, we can't say for sure. At some point, however, excess becomes superfluous. Time will tell.

The 128 and 256GB SM951s raise some questions about the 512GB model's performance. The Lenovo-specific drive is certainly fast, though there are areas we'd expect it to be quicker if there were no restrictions on its speed. The difference in sequential reads between all three models at a queue depth of four is 500 MB/s, or basically a SATA 6Gb/s SSD. 

Samsung's 128GB SM951 is a great value for non-gamers. Games take up a lot of space, and some of us don't get to play as often as we'd like. Others just want a notebook capable of blazing-fast performance. Business users can certainly appreciate the extra battery life, and excellent 4KB read performance makes quick work of large email databases.

The 256GB model gives you room to store a few of your favorite games and professional applications. It delivers 90% of the performance at nearly half the cost of the 512GB drive. That wouldn't mean much in the congested SATA space. But this is PCIe-attached. It's the future, where you'll find the real performance story. Consider 256GB today's sweet spot.

MORE: Best SSDs For The Money
MORE: Latest Storage News

MORE: Storage in the Forums

Chris Ramseyer is a Contributing Editor for Tom's Hardware, covering Storage. Follow him on Twitter and on Facebook.

Follow Tom's Hardware on Twitter, Facebook and Google+.

Create a new thread in the Reviews comments forum about this subject
This thread is closed for comments
43 comments
Comment from the forums
    Your comment
    Top Comments
  • LuxZg
    # Sorry for previous comment, I pressed something * :P * and no edit / delete buddon in view :-/ #

    I'm really hoping for that editorial mentioned in this review. It's unbelivable that last mechanical HDD reviewed by Tom's Hardware was 4TB Seagate in July 2013 !! We've since gone to 8TB drives, new technologies, and so on. I know SSDs are future, but there are still millions of old fashioned HDDs sold each year. I still can't afford a 10TB storage build solely from SSDs.

    So please, do a proper editorial, include everything from Intel 750 and new SM951 NVMe editions, performance SATA SSDs, lower end SATA SSDs, speedy SATA HDDs, large archival "green" HDDs, etc. Keep in mind that people have diverse needs. I want a relatively large SSD in my desktop, but one I can actually afford (not costing more than either CPU or GPU), I want a speedy yet small SSD for my laptop, and can afford more $/GB here because I don't need a 1TB drive, but I want it to have good performance per watt, and I also want large storage drive(s) for desktop, but we're always looking for good performance per $$$ spent. These are widely different categories, and some haven't been covered by Tom's in years.

    Also, when doing the editorial, please keep the charts of "real-world performance" in seconds. It really is a better way to see where to draw the line. If a test lasts 5 minutes and a 1000$ drive shaves 5 seconds over 100$ drive, I don't want to spend extreme $$$$ for those 5 seconds. OK, someone will want to shave even those 5 seconds, but give us info so we can make our buying decisions.

    I personaly am planning on building new PC, AND buying a new laptop, and I really want to make an optimal build(s) where I can decide how many GB/TB I'll appoint in SSDs, and how much in storage HDDs. SSD is getting cheaper all the time, but it's still not THAT cheap if your needs are measured in range of 10TB that you can just go buy some blindly.

    As for last few SSD reviews, they really paint an interesting picture, and even more so makes me ask some questions like above. If you overlap some charts, and you're not a heavy workstation user, it does seem that any SSD drive will do for 99% of tasks.. and saving some serious money and investing it in other subsystems is probably a wise choice, instead of just going after benchmark numbers. I'm way past the "my benchmark is bigger than yours", and if it isn't translated in performance that you can percieve with our limited human senses - it's simply a number on paper, no more.
  • Other Comments
  • Soul_keeper
    Nice review.
    It'd be a shame to see this eventually end up as a 2 horse race intel vs. samsung.
    Here's to hoping marvell and others step up this year.
  • JQB45
    Now we need articles telling us the difference in performance from one motherboard to the next, if any for their built in M2 slots. For example how does an ASUS motherboard compare to say ASRock, etc.
  • hrafn42
    I would think that the biggest "difference in performance" would be between how many, and what generation, PCIe lanes the MB assigns to the M2 slot -- affecting the bandwidth available to the drive. Most "9-Series" MB only toss a couple of 2nd-Gen lanes, resulting in performance not much better than SATA. A few 9-Series MBs peel off some 3rd-Gen lanes from the graphics slot, and most X99 have 3rd Gen lanes to burn so assign a full four of them to the M2.
  • LuxZg
    # Sorry for previous comment, I pressed something * :P * and no edit / delete buddon in view :-/ #

    I'm really hoping for that editorial mentioned in this review. It's unbelivable that last mechanical HDD reviewed by Tom's Hardware was 4TB Seagate in July 2013 !! We've since gone to 8TB drives, new technologies, and so on. I know SSDs are future, but there are still millions of old fashioned HDDs sold each year. I still can't afford a 10TB storage build solely from SSDs.

    So please, do a proper editorial, include everything from Intel 750 and new SM951 NVMe editions, performance SATA SSDs, lower end SATA SSDs, speedy SATA HDDs, large archival "green" HDDs, etc. Keep in mind that people have diverse needs. I want a relatively large SSD in my desktop, but one I can actually afford (not costing more than either CPU or GPU), I want a speedy yet small SSD for my laptop, and can afford more $/GB here because I don't need a 1TB drive, but I want it to have good performance per watt, and I also want large storage drive(s) for desktop, but we're always looking for good performance per $$$ spent. These are widely different categories, and some haven't been covered by Tom's in years.

    Also, when doing the editorial, please keep the charts of "real-world performance" in seconds. It really is a better way to see where to draw the line. If a test lasts 5 minutes and a 1000$ drive shaves 5 seconds over 100$ drive, I don't want to spend extreme $$$$ for those 5 seconds. OK, someone will want to shave even those 5 seconds, but give us info so we can make our buying decisions.

    I personaly am planning on building new PC, AND buying a new laptop, and I really want to make an optimal build(s) where I can decide how many GB/TB I'll appoint in SSDs, and how much in storage HDDs. SSD is getting cheaper all the time, but it's still not THAT cheap if your needs are measured in range of 10TB that you can just go buy some blindly.

    As for last few SSD reviews, they really paint an interesting picture, and even more so makes me ask some questions like above. If you overlap some charts, and you're not a heavy workstation user, it does seem that any SSD drive will do for 99% of tasks.. and saving some serious money and investing it in other subsystems is probably a wise choice, instead of just going after benchmark numbers. I'm way past the "my benchmark is bigger than yours", and if it isn't translated in performance that you can percieve with our limited human senses - it's simply a number on paper, no more.
  • Luay
    Two questions come to mind,

    Do Z97 motherboards with M.2 slots use Sata or PCI-E?

    and, Do video cards installed on a motherboard with an occupied PCI-E M.2 slot operate at x8 bandwidth or at the full x16?
  • danlw
    @LuxZg - I don't see much point in doing performance reviews for mechanical hard drives, other than to show just how much faster SSDs are in comparison. I'm at the point where if I get a mechanical hard drive, it's for archival reasons, and therefore a 5400RPM hard drive is sufficient. It's slower than 7200RPM drives, but I value noise (or rather, lack thereof) over speed when it comes to mechanical storage that I use for archiving. Therefore, the only benchmark metric I care about when it comes to mechanical storage is how quiet (or not) the drive is.

    Ultimately, I'd like to move completely away from mechanical storage. But it will be a few years until a 4TB SSD is affordable. Sure, mechanical storage may retain a $/GB edge, but there comes a point when solid state storage is plentiful and large enough that mechanical hard drives will join tape drives as an archive only format that will fade off into computing history as SSDs become large enough to meet most people's needs as both a system and archival drive.
  • WyomingKnott
    I remember other "interim technologies." Personally, I wouldn't consider getting an m.2 drive until I could have nvme and 4 lanes - they are coming so soon, and the advantage will be so great.

    By the way, I've built my own systems for some time and this is the first time I have ever considered waiting for the next great thing to come out. It's so much more difference than the usual increment.
  • John Philips
    I think it´s no good using it with a Z97 platform, since there are only 16 pcie lanes available, so these ssds may split to 8x for both, which is the case of the Impact vii. Go for X99.
  • Blueberries
    Quote:
    I think it´s no good using it with a Z97 platform, since there are only 16 pcie lanes available, so these ssds may split to 8x for both, which is the case of the Impact vii. Go for X99.


    The SM951 uses 4 lanes, the XP941 uses 2.
  • hrafn42
    Quote:
    The SM951 uses 4 lanes, the XP941 uses 2.

    Actually, the XP941 uses 4 Gen2 lanes, the SM951 4 Gen3 lanes.
  • ko888
    1652322 said:
    I think it´s no good using it with a Z97 platform, since there are only 16 pcie lanes available, so these ssds may split to 8x for both, which is the case of the Impact vii. Go for X99.


    On a Z97 platform the M.2 slot is wired to the Z97 chipset that uses two of the available eight PCI Express 2.0 lanes available from the Z97 chipset itself and has nothing to do with the sixteen PCI Express 3.0 lanes wired to the CPU socket that are dedicated to the graphics card(s) slots.
  • damric
    I want an M.2 SSD so bad now.
  • ykki
    410076 said:
    I want an M.2 SSD so bad now.

    So if you can see they still work after 20 hard resets? :lol:
  • zodiacfml
    The reason why they don't do it in seconds because there's really little difference between these SSDs anymore. Any SSD you could buy today is vastly superior to any HDD.

    Yet, I appreciate these m.2 drives as they are compact versus a 2.5". Performance or battery life doesn't matter much.

    I have seen benchmarks of NVMe drives but the performance means little for client workloads.

    "Also, when doing the editorial, please keep the charts of "real-world performance" in seconds. It really is a better way to see where to draw the line. If a test lasts 5 minutes and a 1000$ drive shaves 5 seconds over 100$ drive, I don't want to spend extreme $$$$ for those 5 seconds. OK, someone will want to shave even those 5 seconds, but give us info so we can make our buying decisions."
  • Filipe Mata
    So the Samsung SM951 has a sequential read Up to 2150MB/s and a sequential Write Up to 1500MB/s. So this means it's 3 times faster than the other ssd's in sequential Write and 4 times faster in sequential read? I've never seen SSD like this.
  • damric
    1427918 said:
    410076 said:
    I want an M.2 SSD so bad now.
    So if you can see they still work after 20 hard resets? :lol:


    50.
  • RedJaron
    414009 said:
    Two questions come to mind, Do Z97 motherboards with M.2 slots use Sata or PCI-E? and, Do video cards installed on a motherboard with an occupied PCI-E M.2 slot operate at x8 bandwidth or at the full x16?

    This isn't a set thing in the chipset. Each motherboard can be configured to use different lanes and different protocols. Thomas usually does a good job in his reviews of mentioning how an M.2 slot is treated.
  • RamCity
    Quote:
    On a Z97 platform the M.2 slot is wired to the Z97 chipset that uses two of the available eight PCI Express 2.0 lanes available from the Z97 chipset itself and has nothing to do with the sixteen PCI Express 3.0 lanes wired to the CPU socket that are dedicated to the graphics card(s) slots.


    Generally that's true. The exception is the ASRock Z97 Extreme6 / Extreme9 which has an 'Ultra M.2' socket with 4x PCIe 3.0 lanes wired directly to the CPU.
  • John Philips
    217811 said:
    1652322 said:
    I think it´s no good using it with a Z97 platform, since there are only 16 pcie lanes available, so these ssds may split to 8x for both, which is the case of the Impact vii. Go for X99.
    On a Z97 platform the M.2 slot is wired to the Z97 chipset that uses two of the available eight PCI Express 2.0 lanes available from the Z97 chipset itself and has nothing to do with the sixteen PCI Express 3.0 lanes wired to the CPU socket that are dedicated to the graphics card(s) slots.


    http://www.asus.com/us/Motherboards/MAXIMUS_VII_IMPACT/specifications/
    Go to the end where you can read "notes". Understand now?
  • John Philips
    1901805 said:
    Quote:
    On a Z97 platform the M.2 slot is wired to the Z97 chipset that uses two of the available eight PCI Express 2.0 lanes available from the Z97 chipset itself and has nothing to do with the sixteen PCI Express 3.0 lanes wired to the CPU socket that are dedicated to the graphics card(s) slots.
    Generally that's true. The exception is the ASRock Z97 Extreme6 / Extreme9 which has an 'Ultra M.2' socket with 4x PCIe 3.0 lanes wired directly to the CPU.


    Where did you see that? So that´s 12 lanes for video and 4 lanes for the ultra m.2 socket?