Flash drives use variations on log-structured storage. The basic thing going on is that the *logical* block numbers being written (which are random) are not the same as the *physical* blocks being written. Drives create 0-N append points, and all those random writes end up becoming sequential writes to pages. At the high end, your write rate can get limited by the page erase rate, which basically translates to an energy/thermal issue (it takes a fair bit of power to erase flash memory pages). The best high end drives can sustain very high mixed read/write rates -- and the key is "sustain" -- for hours/days. Lots of drives out there can handle a short burst of activity for a few tens of seconds, caching everything in RAM on the hardware until the RAM runs out.
Random reads are tougher, because you actually have to go to a random storage block and pull the data. Sequential reads admit lookahead, but random reads don't.
Small writes are cached to DRAM for write combining i.e. multiple IOs are written to NAND at once as the IOs are smaller than the page size. Once the IOs hit the DRAM cache, they are considered complete, hence the higher speed.
And I'm going to make this point again: if even when using NVMe, your random reads are still limited at ~50MB/s, you're only missing on sequential transfers if you stick with AHCI and SATA. Because right now, the bottleneck is elsewhere. Also, for Skylake-U (mobile), SATA offers lower standby power.
50MB/s is a ~50% upgrade over ~30MB/s that SATA offers. It's not even close to what HDD to SSD offers, of course, but we will have to wait for next generation memory for the next huge upgrade.
SATA can do better than ~30MB/s (not sure whether Skylake-U limits the performance in any way, however). NVMe/PCIe still makes sense, because the price premium is not that large. But I'd like to see more reviews highlighting that if you need to save ~$20, going AHCI/SATA is a better option than getting a smaller drive.
I just checked all drives that I benchmarked while at AT. The most was 37MB/s (4KB random read QD1), with most SATA drives hovering between 25-30MB/s. The 256GB SM951-NVMe, on the other hand, did 52MB/s.
While I believe you that the fastest SATA drives that you tested maxed at 30 MBps in random read tests, what makes you think that this is a limitation/bottleneck of the SATA bus rather than the components inside of the SSD (the controller, the NAND flash chips, the number of NAND flash chips, etc)? I do not understand how Ganesh can make statements like "The bandwidth numbers, on the other hand, show that there is a lot to gain by going from SATA to PCIe, and from AHCI to NVMe, and from PCIe 2.0 to PCIe 3.0. However, the gains obtained in each of these upgrades becomes progressively smaller." While I find your benchmark of the 4 SSDs to be very useful, I don't see how you can draw these kinds of conclusions about SATA vs PCIe or AHCI vs NVMe based on your tests.
Also, I don't understand where these 50 MBps and 30 MBps numbers are coming from. Your 850 Pro review from 2014 at http://www.anandtech.com/show/8216/samsung-ssd-850... shows numbers way higher than 30 or 50. Sorry if I'm missing something obvious.
Okay I just re-read the article and think the 50 MBps number is coming from the CrystalDiskMark tests. I think these numbers are very low compared to what other tools will produce, even at QD=1. Regardless, I still don't see how we can draw conclusions about SATA vs PCIe or AHCI vs NVMe based on these results.
I'm totally confused... what are these 30-50MB/s you guys are talking about? Even a crappy USB3 magnetic drive easily does 150MB/s, internal ssds have been pushing the 300MB/s limits of SATA 2 and even 600MB/s of SATA 3 for years now. And before you jump on me with the "random vs sequential" excuse, aren't ssds supposed to make that distinction irrelevant?
I don't know why would you buy a PCI SSD unless you do a lot video work where you really need the best seq speed. For others , even gamers it's 99% waste of money.
i disagree. Load times in multiplayer games are heavily affected by seq speed. Battlefield 4 is a prime example, where fast load times, equal access to the jets and heli's before anyone else.
It's not like you'd round up several graphics cards that produce the same fps, and only run synthetic tests on them to try to show which is fastest. That would be foolish. And that's what's being done with SSDs.
You go as far to claim "those involving heavy multimedia editing and frequent transfers of large-sized files, the PCIe SSDs can definitely provide tangible benefits." How do you know?? If you look at the actual data (in the techreport site), load times of 500-800MB files were pretty much a wash across all the drives. You're misleading readers by only showing PCMark8 and claiming there are tangible benefits.
Dude, did you even read the full article - particularly, the place where the graphs for the 'AnandTech DAS Suite' are displayed? Those graphs are the places where tangible benefit is shown for the PCIe SSDs.
In fact, the only place where I have put in 'synthetic hard drive benchmarks' was the CrystalDiskMark comparisons when introducing the four SSDs. Again, that was prefaced with this text: "...it is useful to determine whether the SSDs are operating as per the manufacturer's claimed specifications. It can also help in finding out whether the SSD is connected via the most optimal interface. ..."
In fact, we set out with this article with the sole intention to use ONLY real-world, application-based benchmarks. Please read the article at least once before putting forward an accusation in its comments section.
It really is a great article, it's well-written, and an interesting read. I'm only focusing on the lack of _real_ real-world benchmarks.
You consider SYSmark, PCMark, and DAS to be real-world, but the problem is they aren't. First, I highly question the accuracy: in an actual load time situation, I seriuosly doubt the 950 Pro will be 5x faster than the Mushkin, as PCMark and DAS show. Secondly, these benchmark programs don't give a tangible understanding: seeing a load time difference in "seconds" is tangible, but seeing scores of a thousand is not.
It's really easy to prove this to yourself: use a stopwatch on anything with a load time. If the times are less than 3x of each other (more likely, within 10%) then it will be evident that PCMark and DAS are lying.
SYSmark is real-world. Look at their whitepaper if you haven't had a chance to try it out. It actually runs the applications and keeps track of how much time it takes to complete tasks - and it actually shows there is little to no difference between a PCIe AHCI SSD operating at 2.0 x4 and a NVMe SSD operating at 3.0 x4.
PCMark - I have linked the PDFs which show how much time it took to complete each workload (real-world trace). The SATA SSD takes around 2 seconds more than the NVMe SSD - and between AHCI and NVMe, it is 0.2 - 0.3s.
The DAS stuff is pretty much as real world as it can be. You have 250GB of data to transfer from one partition to another. The SATA SSD takes 4x the time of the PCIe NVMe SSD. The instantaneous bandwidth numbers are presented in the graph for you to see. Are you saying I am misrepresenting facts?
SYSmark and PCMark are _real_real-world benchmarks - as real-world as you can get if you want highly repeatable benches with reproducible scoring , not something a tech site cooks up on its own (like our DAS suite - which has its own reasons for existence - since we developed it, we can instrument it in ways not possible with third-party benchmarks).
.. in an actual load time situation, I seriuosly doubt the 950 Pro will be 5x faster than the Mushkin, as PCMark and DAS show...
Where does PCMark and DAS say they are representing load time situations? Did you take a look at the PDFs? The PDFs show how much difference is there for the real workload of manipulating images with Photoshop etc. The bandwidth numbers generated by PCMark - I clearly state it is artificial and assumes workload that is not CPU-bound. You should look at the Storage Score to get an idea of how much faster SSD X would be over SSD Y. The bandwidth numbers are only to indicate how the SSDs would perform in a storage-bound situation - Read the explanation preceding the graph.
The DAS suite doesn't talk about load time at all - it notes time taken to transfer a large amount of data from one partition to another. You can see your tangible 'seconds' in those graphs.
Stopwatch and stuff - at the risk of sounding like a broken watch - check the PDFs of the PCMark 8 storage bench results.
I'm not sure why FIO isn't used for your benchmarking. Doesn't have the pretty graphs, but it's got scalability and rigor. You're not going to use Crystal to find out how well a PCIe SSD performs at 600K IOPS, or what happens when you're writing maximum sequential load to 4 of them in a single system.
"What can this hardware do?" and "How will this affect my workload?" are different questions, for sure. I think your application-level benchmarks are quite useful for answering the second. But perhaps not so much for the first.
Or maybe Crystal Diskmark is super-awesome, and FIO's not needed any more. ;)
Your points look technically sound and you clearly have a far better understanding of those benchmark suites than me. The thing is, though, this is confusing. It's not obvious how to take 4 benchmark program results and know how actual computer usage precisely compares.
AT got it right here: http://www.anandtech.com/print/1371/ Load times are easy to comprehend and apply to what we care about! And shockingly, RAID-0 won every suite but marginally lost in the simple use-case test.
I can tell you a lot of people have a misconception that RAID-0 SSDs or the 950 Pro load things >3% faster. [reference to the data: http://www.tomshardware.com/reviews/samsung-950-pr...] The misconception is propagated by reviews that show a bunch of graphs with big performance differences and an omission of simple use-cases. I guess users buying a NUC care about boot time, app load time, etc. Why not show the difference. That would certainly be more meaningful than any *Mark suite.
"We also managed to check out four different SSDs for usage in Skylake-U systems in general (and the NUC6i5SYK in particular)."
The new NUCs are not capable of using the 950 Pro at full speed. Yes, the specs look like they do, but they don't. The maximum speed supported is 1600MB/s. Page 45 of this doc has the specs: http://www.intel.com/content/dam/support/us/en/doc...
Please go through the article at least once before commenting. Did you look at the CrystalDiskMark benchmark numbers with the NEW DEVELOPMENT BIOS? Intel has resolved this issue, and the whole point of this review was going into the reasons behind this.
Yes, notebooks are an interesting category that we have tried to cover in this article, but, without numbers to back up. Our impression is that the lower idle power needed for notebooks might make vendors go with the GT2 OPI setting. So, the numbers are likely to be closer to the BIOS v0042 results that we saw in this review. So, yes, it is likely that PCIe 3.0 x4 SSDs can't get full performance with Skylake-U laptops. Of course, vendors have the knob to turn on the higher performance / higher idle power setting, if they want.
You guys wouldn't happen to have power consumption numbers on the PM951, would you?
It appears the Surface Book's drive is replaceable following a tedious opening procedure, and I've been thinking of replacing the 128 GB PM951 that came with mine with a larger 950 Pro or its successor. The 128 PM drive has terrible slow write performance due to the lack of TurboWrite.
However, I am concerned that the extra performance the 950 and SM series SSDs offer comes at a huge power penalty, and I certainly wouldn't want to give up the fantastic battery life this device offers.
Easiest way is to ask the manufacturer. Otherwise, if the system has a M.2 PCIe x4 slot for a SSD, benchmark something like the SSD 950 PRO using CrystalDiskMark.
Hopefully, after Intel makes this BIOS public, more vendors will put out BIOS updates and spell the OPI link rate out in the change log.
"However, the warranty aspect is a bit worrisome, since the SM951 is an OEM model."
Thank you for mentioning this. I had an SM951 die recently. Newegg's third party seller would not honor the warranty posted (3 years), Newegg would not do anything, and Samsung refused to warranty it. Don't buy the SM951 unless you don't want a warranty and want to risk losing your data.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
40 Comments
Back to Article
sorcio46 - Monday, May 9, 2016 - link
Is there a reason why these flash SSDs have a lower 4K read speed compared with 4K write?James5mith - Monday, May 9, 2016 - link
My guess is writes are buffered vs. reads straight from the raw NAND. But I have no idea if it's actually true.hojnikb - Monday, May 9, 2016 - link
More or less this.rossjudson - Thursday, May 12, 2016 - link
Flash drives use variations on log-structured storage. The basic thing going on is that the *logical* block numbers being written (which are random) are not the same as the *physical* blocks being written. Drives create 0-N append points, and all those random writes end up becoming sequential writes to pages. At the high end, your write rate can get limited by the page erase rate, which basically translates to an energy/thermal issue (it takes a fair bit of power to erase flash memory pages). The best high end drives can sustain very high mixed read/write rates -- and the key is "sustain" -- for hours/days. Lots of drives out there can handle a short burst of activity for a few tens of seconds, caching everything in RAM on the hardware until the RAM runs out.Random reads are tougher, because you actually have to go to a random storage block and pull the data. Sequential reads admit lookahead, but random reads don't.
Kristian Vättö - Monday, May 9, 2016 - link
Small writes are cached to DRAM for write combining i.e. multiple IOs are written to NAND at once as the IOs are smaller than the page size. Once the IOs hit the DRAM cache, they are considered complete, hence the higher speed.dzezik - Thursday, May 12, 2016 - link
SF-2281 has no DRAM. this is not a DRAM cache. the old SandForce 2281 was designed for SLC. the performance with MLC is medicorebug77 - Monday, May 9, 2016 - link
And I'm going to make this point again: if even when using NVMe, your random reads are still limited at ~50MB/s, you're only missing on sequential transfers if you stick with AHCI and SATA. Because right now, the bottleneck is elsewhere.Also, for Skylake-U (mobile), SATA offers lower standby power.
Kristian Vättö - Monday, May 9, 2016 - link
50MB/s is a ~50% upgrade over ~30MB/s that SATA offers. It's not even close to what HDD to SSD offers, of course, but we will have to wait for next generation memory for the next huge upgrade.bug77 - Monday, May 9, 2016 - link
SATA can do better than ~30MB/s (not sure whether Skylake-U limits the performance in any way, however).NVMe/PCIe still makes sense, because the price premium is not that large. But I'd like to see more reviews highlighting that if you need to save ~$20, going AHCI/SATA is a better option than getting a smaller drive.
vladx - Monday, May 9, 2016 - link
Price premium is not large? LMAO it's almost double over SATA ones.Kristian Vättö - Monday, May 9, 2016 - link
I just checked all drives that I benchmarked while at AT. The most was 37MB/s (4KB random read QD1), with most SATA drives hovering between 25-30MB/s. The 256GB SM951-NVMe, on the other hand, did 52MB/s.vladx - Monday, May 9, 2016 - link
Yeah, but paying double for that much is definitely not worth it.Haul22 - Tuesday, May 10, 2016 - link
While I believe you that the fastest SATA drives that you tested maxed at 30 MBps in random read tests, what makes you think that this is a limitation/bottleneck of the SATA bus rather than the components inside of the SSD (the controller, the NAND flash chips, the number of NAND flash chips, etc)? I do not understand how Ganesh can make statements like "The bandwidth numbers, on the other hand, show that there is a lot to gain by going from SATA to PCIe, and from AHCI to NVMe, and from PCIe 2.0 to PCIe 3.0. However, the gains obtained in each of these upgrades becomes progressively smaller." While I find your benchmark of the 4 SSDs to be very useful, I don't see how you can draw these kinds of conclusions about SATA vs PCIe or AHCI vs NVMe based on your tests.Also, I don't understand where these 50 MBps and 30 MBps numbers are coming from. Your 850 Pro review from 2014 at http://www.anandtech.com/show/8216/samsung-ssd-850... shows numbers way higher than 30 or 50. Sorry if I'm missing something obvious.
Haul22 - Tuesday, May 10, 2016 - link
Okay I just re-read the article and think the 50 MBps number is coming from the CrystalDiskMark tests. I think these numbers are very low compared to what other tools will produce, even at QD=1. Regardless, I still don't see how we can draw conclusions about SATA vs PCIe or AHCI vs NVMe based on these results.Visual - Thursday, May 12, 2016 - link
I'm totally confused... what are these 30-50MB/s you guys are talking about? Even a crappy USB3 magnetic drive easily does 150MB/s, internal ssds have been pushing the 300MB/s limits of SATA 2 and even 600MB/s of SATA 3 for years now. And before you jump on me with the "random vs sequential" excuse, aren't ssds supposed to make that distinction irrelevant?TheinsanegamerN - Monday, May 9, 2016 - link
The premium is pretty high. A 256GB 950 pro m.2 is $189, a 1TB sata ssd is $209. That is a pretty big difference.Heck, last year I got a 512GB SSD for less than $189.
dzezik - Thursday, May 12, 2016 - link
this is limited by very old sandforce controller not designed for MLC but for SLCvladx - Monday, May 9, 2016 - link
I don't know why would you buy a PCI SSD unless you do a lot video work where you really need the best seq speed. For others , even gamers it's 99% waste of money.jasonelmore - Monday, May 9, 2016 - link
i disagree. Load times in multiplayer games are heavily affected by seq speed. Battlefield 4 is a prime example, where fast load times, equal access to the jets and heli's before anyone else.A5 - Monday, May 9, 2016 - link
Not if you play on a competently admined server. This is what pre-round timers are for.AnnonymousCoward - Monday, May 9, 2016 - link
When will AnandTech realize that synthetic hard drive benchmarks are utterly pointless?http://techreport.com/review/29221/samsung-950-pro...
It's not like you'd round up several graphics cards that produce the same fps, and only run synthetic tests on them to try to show which is fastest. That would be foolish. And that's what's being done with SSDs.
You go as far to claim "those involving heavy multimedia editing and frequent transfers of large-sized files, the PCIe SSDs can definitely provide tangible benefits." How do you know?? If you look at the actual data (in the techreport site), load times of 500-800MB files were pretty much a wash across all the drives. You're misleading readers by only showing PCMark8 and claiming there are tangible benefits.
ganeshts - Monday, May 9, 2016 - link
Dude, did you even read the full article - particularly, the place where the graphs for the 'AnandTech DAS Suite' are displayed? Those graphs are the places where tangible benefit is shown for the PCIe SSDs.In fact, the only place where I have put in 'synthetic hard drive benchmarks' was the CrystalDiskMark comparisons when introducing the four SSDs. Again, that was prefaced with this text: "...it is useful to determine whether the SSDs are operating as per the manufacturer's claimed specifications. It can also help in finding out whether the SSD is connected via the most optimal interface. ..."
In fact, we set out with this article with the sole intention to use ONLY real-world, application-based benchmarks. Please read the article at least once before putting forward an accusation in its comments section.
AnnonymousCoward - Monday, May 9, 2016 - link
It really is a great article, it's well-written, and an interesting read. I'm only focusing on the lack of _real_ real-world benchmarks.You consider SYSmark, PCMark, and DAS to be real-world, but the problem is they aren't. First, I highly question the accuracy: in an actual load time situation, I seriuosly doubt the 950 Pro will be 5x faster than the Mushkin, as PCMark and DAS show. Secondly, these benchmark programs don't give a tangible understanding: seeing a load time difference in "seconds" is tangible, but seeing scores of a thousand is not.
It's really easy to prove this to yourself: use a stopwatch on anything with a load time. If the times are less than 3x of each other (more likely, within 10%) then it will be evident that PCMark and DAS are lying.
ganeshts - Tuesday, May 10, 2016 - link
SYSmark is real-world. Look at their whitepaper if you haven't had a chance to try it out. It actually runs the applications and keeps track of how much time it takes to complete tasks - and it actually shows there is little to no difference between a PCIe AHCI SSD operating at 2.0 x4 and a NVMe SSD operating at 3.0 x4.PCMark - I have linked the PDFs which show how much time it took to complete each workload (real-world trace). The SATA SSD takes around 2 seconds more than the NVMe SSD - and between AHCI and NVMe, it is 0.2 - 0.3s.
The DAS stuff is pretty much as real world as it can be. You have 250GB of data to transfer from one partition to another. The SATA SSD takes 4x the time of the PCIe NVMe SSD. The instantaneous bandwidth numbers are presented in the graph for you to see. Are you saying I am misrepresenting facts?
SYSmark and PCMark are _real_real-world benchmarks - as real-world as you can get if you want highly repeatable benches with reproducible scoring , not something a tech site cooks up on its own (like our DAS suite - which has its own reasons for existence - since we developed it, we can instrument it in ways not possible with third-party benchmarks).
Where does PCMark and DAS say they are representing load time situations? Did you take a look at the PDFs? The PDFs show how much difference is there for the real workload of manipulating images with Photoshop etc. The bandwidth numbers generated by PCMark - I clearly state it is artificial and assumes workload that is not CPU-bound. You should look at the Storage Score to get an idea of how much faster SSD X would be over SSD Y. The bandwidth numbers are only to indicate how the SSDs would perform in a storage-bound situation - Read the explanation preceding the graph.
The DAS suite doesn't talk about load time at all - it notes time taken to transfer a large amount of data from one partition to another. You can see your tangible 'seconds' in those graphs.
Stopwatch and stuff - at the risk of sounding like a broken watch - check the PDFs of the PCMark 8 storage bench results.
rossjudson - Tuesday, May 10, 2016 - link
I'm not sure why FIO isn't used for your benchmarking. Doesn't have the pretty graphs, but it's got scalability and rigor. You're not going to use Crystal to find out how well a PCIe SSD performs at 600K IOPS, or what happens when you're writing maximum sequential load to 4 of them in a single system."What can this hardware do?" and "How will this affect my workload?" are different questions, for sure. I think your application-level benchmarks are quite useful for answering the second. But perhaps not so much for the first.
Or maybe Crystal Diskmark is super-awesome, and FIO's not needed any more. ;)
AnnonymousCoward - Wednesday, May 11, 2016 - link
Your points look technically sound and you clearly have a far better understanding of those benchmark suites than me. The thing is, though, this is confusing. It's not obvious how to take 4 benchmark program results and know how actual computer usage precisely compares.AT got it right here: http://www.anandtech.com/print/1371/ Load times are easy to comprehend and apply to what we care about! And shockingly, RAID-0 won every suite but marginally lost in the simple use-case test.
I can tell you a lot of people have a misconception that RAID-0 SSDs or the 950 Pro load things >3% faster. [reference to the data: http://www.tomshardware.com/reviews/samsung-950-pr...] The misconception is propagated by reviews that show a bunch of graphs with big performance differences and an omission of simple use-cases. I guess users buying a NUC care about boot time, app load time, etc. Why not show the difference. That would certainly be more meaningful than any *Mark suite.
Agent Smith - Monday, May 9, 2016 - link
Not so quick to apologise eah?Experientia docet
Kvaern2 - Monday, May 9, 2016 - link
Much can be read in a username.MrSpadge - Monday, May 9, 2016 - link
"Please read the article at least once before putting forward an accusation in its comments section."Non, no! That's not how the internets are supposed to work ;)
MrSpadge - Monday, May 9, 2016 - link
(Oops, meant to reply to Ganesh's post)BPB - Monday, May 9, 2016 - link
"We also managed to check out four different SSDs for usage in Skylake-U systems in general (and the NUC6i5SYK in particular)."The new NUCs are not capable of using the 950 Pro at full speed. Yes, the specs look like they do, but they don't. The maximum speed supported is 1600MB/s. Page 45 of this doc has the specs:
http://www.intel.com/content/dam/support/us/en/doc...
You can also view a discussion at the Intel Community regarding this:
https://communities.intel.com/message/372201#37220...
This being the case, which SSD makes the best buy?
ganeshts - Monday, May 9, 2016 - link
Please go through the article at least once before commenting. Did you look at the CrystalDiskMark benchmark numbers with the NEW DEVELOPMENT BIOS? Intel has resolved this issue, and the whole point of this review was going into the reasons behind this.BPB - Monday, May 9, 2016 - link
Sorry, read while at work and only had time to breeze through. I looked for those numbers but didn't see them. Wish I could delete my post...willis936 - Monday, May 9, 2016 - link
Does this apply to all -U processors? I think far more interesting than NUCs are people with broadwell and skylake -U laptops.ganeshts - Monday, May 9, 2016 - link
This is applicable to all Skylake-U processors.Yes, notebooks are an interesting category that we have tried to cover in this article, but, without numbers to back up. Our impression is that the lower idle power needed for notebooks might make vendors go with the GT2 OPI setting. So, the numbers are likely to be closer to the BIOS v0042 results that we saw in this review. So, yes, it is likely that PCIe 3.0 x4 SSDs can't get full performance with Skylake-U laptops. Of course, vendors have the knob to turn on the higher performance / higher idle power setting, if they want.
rish95 - Tuesday, May 10, 2016 - link
You guys wouldn't happen to have power consumption numbers on the PM951, would you?It appears the Surface Book's drive is replaceable following a tedious opening procedure, and I've been thinking of replacing the 128 GB PM951 that came with mine with a larger 950 Pro or its successor. The 128 PM drive has terrible slow write performance due to the lack of TurboWrite.
However, I am concerned that the extra performance the 950 and SM series SSDs offer comes at a huge power penalty, and I certainly wouldn't want to give up the fantastic battery life this device offers.
ez76 - Monday, May 9, 2016 - link
How can I tell if my Skylake-U system has an OPT GT2 or OPT GT4 configuration?ganeshts - Monday, May 9, 2016 - link
Easiest way is to ask the manufacturer. Otherwise, if the system has a M.2 PCIe x4 slot for a SSD, benchmark something like the SSD 950 PRO using CrystalDiskMark.Hopefully, after Intel makes this BIOS public, more vendors will put out BIOS updates and spell the OPI link rate out in the change log.
Junkmail1 - Monday, May 9, 2016 - link
I went through the article twice and the comments once. Now I'm second guessing that choice!buyme50 - Saturday, May 14, 2016 - link
"However, the warranty aspect is a bit worrisome, since the SM951 is an OEM model."Thank you for mentioning this. I had an SM951 die recently. Newegg's third party seller would not honor the warranty posted (3 years), Newegg would not do anything, and Samsung refused to warranty it. Don't buy the SM951 unless you don't want a warranty and want to risk losing your data.