Samsung SM951-NVMe Versus AHCI Versus SATA 850 Pro

Indroduction

Several readers commented on the lack of 2.5" SATA drives in our SM951 coverage. Comparing SATA to PCIe-based storage isn't apples-to-apples. I tend to make comparisons based on interface technology or price tag. Samsung's 850 Pro is a very fast 2.5" SATA SSD, but it's not in the same league as the company's SM951.

Don't believe me? In this round-up of Samsung products, I'm pitting the 850 Pro 256GB against Samsung's SM951 AHCI 256GB. I'll also introduce you to the SM951-NVMe 256GB, which should be available in June.

At this time, only the 256GB model is available from Samsung's SM951-NVMe product family. Some sites published reviews of the SM951-NVMe already using drives harvested from Intel NUCs. But we've learned that Intel's samples didn't ship with final releases firmware. As a result, we shelved our early review using one of those SSDs borrowed from a NUC.

Intel wasn't the only company to receive early SM951-NVMe SSDs. At least one distributor had products in-hand, ready to ship before Samsung asked for them back to update the firmware. If that firmware wasn't required, the SM951-NVMe would already be on sale. 

Technical Specifications

The AHCI-based SM951 doesn't actually have AHCI in its product name. The new NVMe-capable drive does, though (SM951-NVMe). Keep that in mind when you place your order.

MORE: Best SSDs For The Money
MORE: How We Test HDDs And SSDs

MORE: All Storage Content

As you can see, the PCIe-based SM951s are up to three times faster than the SATA 6Gb/s-attached 850 Pros in certain categories. We've already covered the Samsung 850 Pro in detail here, and the SM951 AHCI here and here. With that said, let's focus on the SM951-NVMe.

Faced with sequential data, the SM951-NVMe is only a little faster than the AHCI models. The real performance increase comes from reading random data. The SM951 (AHCI) spec sheet claims up to 90,000 random read IOPS. That's less than the 850 Pro, which tops out at up to 100,000 IOPS. Although we managed to pull more random read IOPS from the AHCI drive, it couldn't come close to the SM951-NVMe's claimed 300,000 random read IOPS. 

Random write IOPS also improve. The 256 and 512GB SM951-NVMes deliver up to 100,000 IOPS. The smaller 128GB model peaks at 83,000 IOPS. Those numbers are up from the 70,0000 IOPS figure that Samsung ascribes to the SM951 AHCI. 

We believe Samsung stopped measuring IOPS performance at queue depth of 32 (SATA's limit). Why do we think this? At higher queue depths, we measured the SM951 AHCI in excess of 160,000 random read IOPS. And using a single worker, we were not able to match Samsung's SM951-NVMe spec of 300,000 IOPS.

Recently, we published data showing the SM951 AHCI we pulled from a Lenovo Ultrabook was slower than the same drive purchased from RamCity. Lenovo has since updated its firmware, though, bringing performance up to the expected levels. You can find that firmware here for SM951 SSDs purchased in Lenovo notebooks.

Pricing And Accessories

The SM951-NVMe still isn't available to buy, but we expect that to change in the coming weeks. We reached out to Rod Bland at RamCity in Australia for comment. He responded, "Initial expectations were for stock to be available to SSI channel partners just a couple of weeks from now, but Samsung wanted to wait until a new firmware update was distributed first. That has delayed shipping of product in all markets, so we are back to  around mid-July before we expect to have stock of the NVMe SM951.”

Since these are OEM parts, it's up to resellers to package the drives for delivery and to customize the accessory package. Most of the drives we've tested ship in an anti-static bag inside a foam-lined box. Unfortunately, the SSDs from Samsung's SSI group do not work with the company's Magician software.

A Closer Look

Originally I wasn't going to remove the stickers on the SM951 and SM951-NVMe, since both parts should be identical. But as I took pictures, I found that they're aren't exactly the same, though.

In the shot above, the NVMe model is on the top; you can identify it by the third letter in the part number. The NVMe model is designated by a V, while the AHCI model has an H.

The same Samsung three-core UBX controller is used on both SSDs.

The same low-power DDR2 DRAM is also used on both 256GB parts.

We took a picture of the back of the drives to show that this capacity size is single-sided. But in doing so, we noticed that one drive was a little taller. After investigating further, we discovered different flash. The SM951 AHCI uses an eight-die stack (per Samsung's NAND flash decoder) and the SM951-NVMe uses a 16-die stack.

For the most part, though, the flash is the same other than its die density and the number of dies in each package.

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.

The Samsung 850 Pro is a great SSD, even if the SATA 6Gb/s interface limits its performance. We're confident that 3D V-NAND is capable of delivering higher performance than Samsung's two-bit MLC NAND. However, the controller isn't capable of expressing what's theoretically possible on that bottlenecked interface.

The two SM951 drives break free from the SATA standard introduced all the way back in 2008. In the two charts above, we see that even PCIe-based storage using AHCI, a command protocol introduced with Windows Vista, is a crutch to flash storage devices. Peak sequential read performance is identical between the two SM951 256GB SSDs, but we observed a large divide at low queue depths. NVMe has lower latency than AHCI, and that increases performance when there aren't many outstanding commands.

Most desktop workloads fall into the low queue depth category, and I'd rather see a small speed-up at low queue depths over a large increase at high queue depths. That's usable performance. Perhaps that justifies the SM951-NVMe's added expense, but what does it all mean for Samsung's 850 Pro?

This is why we rarely mix SATA-attached drives into our PCIe-based comparisons. The performance gap is very wide, like comparing hard drives to SSDs. When solid-state storage first emerged, readers wanted to see us compare the technology to the fastest mechanical disks. We are still in a place where enthusiasts need to see the difference with SSDs, even though the gap is massive. Next-generation isn't just a marketing term for the companies manufacturing these parts.

Sequential Write

Both SM951s generate nearly the same performance in our sequential write test. Here, the bottleneck is the flash, and not the protocol. Many hope that Samsung introduces a PCIe-based NVMe client SSD with 3D V-NAND at its SSD Global Summit, an annual gathering for the press it usually puts on in July. If that product comes to pass, we will finally get to compare Samsung's 1xnm MLC to 3D V-NAND without SATA interfering.

Random Read

Since NVMe helps reduce latency, it shouldn't come as a surprise to see the largest gains coming from random access patterns. Again, Samsung's 1xnm MLC flash is fast enough while reading data to demonstrate the benefit of NVMe. The SM951-NVMe 256GB SSD reaches just over 250,000 IOPS before leveling off. With the same controller and flash, but the older AHCI protocol, the SM951 256GB is down by nearly 100,000 peak random read IOPS.

Of course, peak performance is interesting. But most of us will notice the additional performance at lower queue depths. At QD1, the NVMe model delivers around 2000 more random read IOPS, which is equivalent to a typical SSD product cycle speed-up. Those usually happen each year. But Samsung is delivering a generational upgrade in just a few months. 

Comparing the SM951-NVMe to the 850 Pro, we see the equivalent of a two-generation increase at low queue depths. Most users spend 80% of their time reading data and around 20% writing it. Those requests are typically random in nature. Given both stipulations, you're looking at a very large performance boost over the 850 Pro.

We were able to reach Samsung's claimed 300,000 random read IOPS by changing the parameters of our test. In order to do so, we had to mimic enterprise-class benchmarking, veering away from our consumer workloads.

Random Write

Our SSD test procedure includes a limited form of preconditioning. The drives are not taken down to true steady state, which is something we do for our enterprise SSD reviews on Tom's IT Pro. But the tests are not performed on drives fresh out of the box, either.

Samsung claims 90,000 random write IOPS on both SM951s, and we achieved that mark fairly easily. Fresh out of the box, the number rises significantly, too.

According to Samsung, the 256GB 850 Pro can reach up to 100,000 random write IOPS. That number is realistic, though not after our preconditioning process, where garbage collection and wear-leveling algorithms come into play. Every flash controller company handles these activities differently. Later in this review, we'll detail the results of how Samsung cleans house in the background.

80% Read Sequential Mixed Workload

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

The SM951-NVMe delivers higher sequential read performance than the AHCI model at low queue depths when it's only reading data. With mixed reads and writes, the SM951-NVMe maintains a slight lead at QD1, though the gap shrinks. At QD2, performance is nearly identical. Everything above that queue depth favors the AHCI model.

We aren't sure why the NVMe model doesn't show a strong lead in this test, except that perhaps Samsung's SM951 architecture isn't yet fully optimized.

80% Read Random Mixed Workload

Conversely, with random mixed read and write data, the SM951-NVMe scales better with queue depth than the AHCI model. At low queue depths, the performance difference is much more subtle. You wouldn't notice it under real-world conditions.

Sequential Steady State

The two 256GB SM951s deliver nearly identical steady state sequential performance. The same is true for random 4KB steady state. This test stresses architecture and programing more than the physical flash or controller. 

The area we want to really focus on is the performance increase over Samsung's 850 Pro 256GB. We often see reader comments that downplay the delta between SATA and PCIe. It's true that, in some applications, one won't be drastically better than the other. It all comes down to the workloads you subject your storage to.

After a reasonable number of sequential and random writes, we tested the SM951-NVMe with HD Tach to watch sequential performance drop off and background activity kick in. Controller makers use proprietary code to handle those tasks. Sometimes the code changes from firmware updates. But most of the time, alterations are implemented with new processor launches.

Samsung's client SSDs are not very aggressive when it comes to TRIM and cleaning dirty cells. Under heavy real-world workloads, we often see performance drop off. All client SSDs transition from fresh-out-of-box to steady state eventually. But the amount of work it takes to get there and the amount of time it takes for the drive to recover is what we're looking for. Samsung's SSDs in particular tend to fall off faster and take longer to recover than other products.

That isn't always a bad thing, though. It means Samsung SSDs wear their flash less because of lower write amplification. In order to clean cells for fresh writes, the entire page of flash needs to be read and the data rewritten without old data that was deleted. Some SSDs are very aggressive with this process, causing the flash to fail faster. Also, the read, erase and write cycle uses power, so aggressive background activity means higher consumption over time. This affects notebook battery life a great deal.

Random Write Steady State

With random 4KB steady state writes, we're looking at two things. The obvious first observation is IOPS performance, where a higher number is better. The second variable is how closely bound the minimum and maximum IOPS are in relation to one another. This is what we call performance consistency.

The two 256GB SM951s deliver higher peak random write steady state IOPS than the 850 Pro, but the deviation between low and high marks is larger. Some enterprise products will actually limit random write IOPS in order to deliver a more consistent flow of data for applications that rely on steady performance over peak performance.

PCMark 8 Real-World Software

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

Here's your first real look at the two SM951s compared to Samsung's 850 Pro in real-world desktop applications. We didn't expect to see the SM951-NVMe improve over the AHCI model, but that's not the case; there is in fact a slight speed-up. Many of these tests run in a single thread at low queue depths, the Photoshop Heavy metric serving as the sole exception. 

Singling out that benchmark, the two SM951s finish within two seconds of each other. The SM951-NVMe wraps up the job a full 10 seconds faster than Samsung's 850 Pro. Photoshop Heavy also shows the widest gap between the three drives. Most of the other benchmarks are separated by a few seconds or less.

Each run takes about 20 minutes to complete. Over that span of time, we see a large divide between the three products in question. The results above are presented in MB/s rather than time to complete, if only to show that those small time differences really do equate to a big impact on user experience.

PCMark 8 Advanced Workload

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

The PCMark 8 workload tests use the same data as PCMark 8's real-world benchmarks, but run the modules back to back after preconditioning the drives. After steady state is achieved, time is added between runs to allow the wear-leveling algorithms to recover performance. 

We've already mentioned that Samsung SSDs are not as aggressive as some others, so the clean-up time takes longer. Since we're only comparing Samsung products, we don't need to compare the company's wear-leveling to its competition. We will say, however, that Samsung's drives would deliver higher performance in the recovery phase of this test if the five-minute pause between runs was longer. Performance would also improve if Samsung over-provisioned its SSDs. As they sit from the factory, all three drives give you the full LBA span to store data. Of course, you're able to configure these 256GB models with 240GB of user space and 16GB allocated for background activities during the format process.

The throughput results show what we expected after seeing the PCMark 8 real-world results. Again, these three drives behave nearly the same under identical load levels due to their underlying architecture. The difference in performance is clear, though.

The service times show us how long each test iteration takes to complete all of the tasks. The 256GB SM951s deviate less than the 850 Pro 256GB when measuring heavy and light workloads. Most desktop users should focus more on the recovery portion for an indication of daily use performance though, while professional users running write-intensive workstation applications should focus on the middle steady state area.

The SM951 products were designed to provide superior performance for professional users. What makes them so attractive for everyone is their low latency during the recovery phase. Samsung didn't really design the SM951 for light workloads; its success there is just a byproduct of making very good workstation SSDs.

Conclusion

Did you notice the missing Notebook Battery Life tests that we normally run on client SSDs? This brings us to a sour note when dealing with NVMe: compatibility. Even though our Lenovo X1 Carbon Gen 3 fully supports the AHCI-attached SM951, the NVMe model will not boot in the Broadwell-based platform. Lenovo may add support at a later date, but as of right now, the system kicks out an error telling us that it can't start up. Without boot support, we can't run our barrage of battery life tests on the SM951-NVMe.

And this points to an even larger problem. Several motherboard manufacturers released NVMe-enabled firmware updates for new X99 and Z97 platforms. But older boards haven't received the same love. Many users upgrading to cutting-edge storage products are building them into completely new configurations. There is an even larger group of enthusiasts with older Z87- and Z77-based machines, though. If your board vendor of choice isn't deliberately adding NVMe support, you won't be booting to an NVMe drive any time soon. We hope this will happen over time, but we aren't holding our breath that years-old motherboards will suddenly get new features.

If you already purchased a Samsung SM951 SSD with AHCI, then we can't recommend the NVMe model for its slight performance boost. It's faster, but only slightly. Moving from an 850 Pro to the SM951-NVMe is a nice upgrade though, and you will notice snappier performance from the reduced latency. 

Compatibility obviously needs to be considered, but we also feel capacity should be included in your purchasing decision. The SM951s are limited to a maximum capacity of 512GB, while Samsung's 850 Pro scales to 1TB. That could be a problem for you if your Steam/Origin folder is bursting at the seams.

The 850 Pro 256GB is one of the fastest SATA-based products on the market today. Samsung's SM951-NVMe 256GB is the fastest client PCIe drive smaller than 400GB. We expect it to sell for just over $256 when it's released. That's much less than Intel's SSD 750 400GB NVMe SSD. If you want performance beyond what SATA offers but don't want to spend big money on unused capacity, the smaller SM951-NVMes are a good choice. 

When they become available, we'll published performance results from the SM951-NVMe 128GB and 512GB models. We're excited to see how the SM951-NVMe 512GB compares to Intel's SSD 750 400GB.

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 Facebook.

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

Create a new thread in the US Reviews comments forum about this subject
This thread is closed for comments
24 comments
    Your comment
  • killabanks
    Glad to see that some aspects of computing are getting rapid improvements.
    3
  • Amdlova
    I will stick with 2.5" For Heat problems lol
    0
  • Amdlova
    hpv-2560 is rated 1.9A and the another one VPV-2560 is 2.7A Maybe you guys need test the power consumption.
    0
  • atheus
    I find the preamble about comparing these M.2 drives to regular SATA drives a little odd. When SSD's first came out, of course they were compared to HDD's because that was precisely the question of the moment — how much more performance can one expect from the more expensive SSD's. With this new generation of SSD's, it's only natural to want to see them compared to the old, cheaper stuff. How else are we going to evaluate whether it's worth it to upgrade?

    If it only resulted in an extra 10% on a real-world test, then it wouldn't make any sense to spend double or more on a NVMe drive. With these solid numbers, though, the massive performance leap will be well worth the cost for those who can afford it.
    2
  • TechyInAZ
    Looks great! NVMe is defiantly the ideal solution for users that need high speed/responsive storage.

    I wish you guys would of compared this NVMe SSD to the Intel NVMe SSD that just came out.
    1
  • CRamseyer
    The Intel SSD 750 and the Samsung SM951-NVMe will show up side by side soon in another review.
    2
  • barnyrubble
    What about the rest of the system in which they were tested?
    0
  • CRamseyer
    Test systems: http://www.tomshardware.com/reviews/how-we-test-storage,4058.html

    All three drives were tested in the PCIe test system. They are the only systems setup for testing queue depths beyond 32.
    1
  • dwnelson
    I'm sad the real world chart is misleading. If you show seconds instead of MB/s you'd see it doesn't justify the cost. Sad.
    1
  • dark_wizzie
    Quote:
    The Intel SSD 750 and the Samsung SM951-NVMe will show up side by side soon in another review.

    That is a natural comparison and I look forward to reading it.
    0
  • philipemaciel
    So going from a SATA SSD to a Pci-e SSD brings huge improvements in benchmarks, but only incremental gains in real life applications?
    2
  • CRamseyer
    Anonymous said:
    So going from a SATA SSD to a Pci-e SSD brings huge improvements in benchmarks, but only incremental gains in real life applications?


    In single applications it is just another incremental performance increase in 'time to complete' tasks. You do get a nice reduction in latency, 'time to start'. The latter improves your user experience.

    For workstation users running professional applications performance increases all across the board. In a workstation environment the queue depth loads increase so those users can take advantage of the increased performance at higher queue depths.
    2
  • Bayfront Benny
    There are other backward compatibility issues with NVMe. To my knowledge you can not clone these drives with existing cloning software (I couldn't clone the Intel 750 with Acronis), which does mean that if you want to try before you buy you have to do a complete install. Install has to be from UEFI USB I believe a per Rod's blog at RamCity. I don't think you can use DVD. Correct me if I'm wrong though. This install is OK - I've done it but definitely not something you want to do often if you really want to install all your programs and system the way you like.
    0
  • CRamseyer
    I didn't have an issue cloning to the SM951-NVMe with the latest version of Acronis. I haven't tried to clone to an Intel SSD 750.
    0
  • mapesdhs
    (with respect to Photoshop Service Time...)

    "...a full 10 seconds faster..."

    That sounds like a lot because we can all imagine the notion of waiting for 10 seconds for anything, but it would have been clearer and more sensible to say the NVMe was just under 3% faster, and that's really not very much at all. Indeed, in relative terms those charts show minimal speedups, to a degree that IMO means any spare cash would be much better spent on some other part of the system for those apps, or in some cases the app in question can benefit from an SSD but not in a manner that is tested here.

    For example, After Effects runs best with an SSD used to hold the main and media caches, so it's an obvious thing to test for any gain with a PCIe SSD, and again NMVe over an AHCI PCI solution. You could try a render or other task, since any improvement in interactive response or render time is what matters, but it could be difficult to determine the former.

    The Service Time charts don't look at all useful to me. The gains with the NVMe in each case could be a lot more, but one cannot tell as the test isn't specific to where having an SSD can make a difference for each app, AE especially.

    Ian.
    1
  • Eggz
    If this is just the first generation of consumer NVMe, then that's pretty exciting! Especially considering that it's using older flash. I can't wait to see what's out when it's time for me to either upgrade or build another system.
    1
  • Bill_VA1
    Great article. I'm assuming that the SM951-NVMe would still be faster than 2 Samsung SATA3 Pros in a RAID0?
    0
  • Eggz
    Anonymous said:
    Great article. I'm assuming that the SM951-NVMe would still be faster than 2 Samsung SATA3 Pros in a RAID0?


    Yeah, where it matters (i.e. random workload), it definitely would. Doing RAID 0 doesn't make the drives retrieve small data bits any faster, just sequential operations. That is, unless you start putting together a lot of drives in RAID 0, which starts to scale the 4K random operations because data bits are striped. I don't think just two of them would make much of a difference, though. Even if it did, two 950 Pros wouldn't catch up to the 950 NVMe.

    http://www.tomshardware.com/reviews/ssd-dc-s3700-raid-0-benchmarks,3480-4.html
    1
  • MittRomney4President
    Will it support hardware based encryption???
    2
  • Phuntasm
    Seems like high bandwidth, but the real world effects are minimal. None of those benchmarks aside for artificial ones showed any end-user perceivable gains. Marginal increase at best. Given the price difference it seems this money is better spent elsewhere
    1