Review Index:
Feedback

Samsung 960 PRO 2TB M.2 NVMe SSD Full Review - Even Faster!

Subject: Storage
Manufacturer: Samsung

Introduction, Specifications and Packaging

Introduction:

Just under a year ago we published our review of the Samsung 950 PRO, their first foray into NVMe SSD territory. Today we have a 960 PRO, which strives to be more revolutionary than evolutionary. There are some neat new features like 16-die packages and a Package-on-Package controller/DRAM design, all cooled by a copper heat spreading label! This new model promises to achieve some very impressive results, so without further delay, let's get to it!

View Full Size

Specifications:

View Full Size

View Full Size

Specs have not changed since the announcement. Highlights include

  • A new 5-core Polaris controller (with one die solely dedicated to coordinating IO's to/from the host)
  • 4-Landing Design - It's tough fitting four flash packages onto an M.2 2280 SSD, but Samsung has done it, thanks to the below feature.
  • Package-on-Package - The controller and DRAM are stacked within the same package, saving space.
  • Hexadecimal Die Packages - For the 960 Pro to reach 2TB of capacity, 16 48-layer MLC V-NAND packages must be present within each package. That's a lot of dies per package!

Packaging:

View Full Size

View Full Size

Nice touch with the felt pad on the bottom of the installation guide. This pad keeps the 960 Pro safely in place during shipment.

Read on for the full review of the 2TB Samsung 960 PRO!


October 18, 2016 | 11:25 AM - Posted by Biased (not verified)

It looks like we are getting very close to the theoretical limit of the M.2 interface. What's going to happen after that. Add in cards on x8 on the consumer side or are we just going to see stagnation for the next few years like we have with SATA 3.0 SSDs

October 18, 2016 | 11:43 AM - Posted by Allyn Malventano

Higher capacities, faster flash and controllers. If you can set aside the megapixel race that has been going on to achieve 'max' figures, the new race is going to be achieving higher throughputs at lower queue depths. If you check out the QD Weighted results, you'll see there is still plenty of room to go before saturating PCIe 3.0 x4.

October 18, 2016 | 03:13 PM - Posted by John H (not verified)

Hey Allyn - do you have any way to calculate what the max IOPs are for a given bandwidth?

I assume it's something like 100K IOPs per ~ 500MB/sec based upon fully saturated 6gbps SATA SSDs? Maybe 2-4x higher as NVMe has lower overhead?

I'm just trying to figure out how much improvement is possible over the QD=1 results of this Samsung 960 Pro on a full x4 PCI-E link..

Thanks!

October 18, 2016 | 11:02 PM - Posted by Allyn Malventano

3.6 GBps / 4KB = 900,000 IOps. This figure was also seen in the Micron QuantX demo.

There are other limiting factors (per-IO latency across the bus, etc) that will limit QD=1 to a bit lower than complete saturation, but we still have a way to go.

October 18, 2016 | 12:35 PM - Posted by Paul A. Mitchell (not verified)

> It looks like we are getting very close to the theoretical limit of the M.2 interface. What's going to happen after that?

Sort term: NVMe RAID controllers like the Highpoint RocketRAID 3840A with full x16 PCIe 3.0 edge connectors.

(I've been scouring the Internet for news of similar cards, but the 3840A is the only one with published specs.)

Some x16 cards support four NVMe ports, but they are
"pass-thru" designs that only support software RAID arrays
and, as such they are not bootable.

Some x16 AOCs support four M.2 ports on-board,
e.g. Dell, HP, Kingston, the "Squid": but those
are also pass-thru designs that can only boot
from a single M.2 (no RAID support).

Intel's DMI 3.0 link has the same upstream bandwidth
as a single NVMe M.2 port, so there's little to be gained from a RAID-0 array that is downstream of that DMI 3.0 link:
possibly an increase in average WRITE speeds.

Medium term: more NVMe RAID controllers, with either x8 or x16 edge connectors. The x16 edge connector is more "elegant" because it maps neatly to four x4 lanes.

Long term:
(a) PCIe 4.0's clock will be 16 GHz;
(b) Intel may widen future DMI links with more lanes;
(c) we're awaiting specs for AMD's Zen chipsets.

Hope this helps.

October 22, 2016 | 02:11 PM - Posted by Anonymouse (not verified)

Here's to 40 PCIe lanes off HEDT cpu then!

October 30, 2016 | 06:31 PM - Posted by Logic

IIRC the DMI bus and the PCIe bus are locked together, so OCing the PCIe bus should be another short term solution to raising the DMI bottleneck.

IIRC/IMHO some components use the PCIe frequency as a base frequency and run their chip frequencies at a multiple of it.
Thats what is probably causing the horror stories around PCIe overclocking...

With benckmarks, from:
http://www.overclock.net/t/1461359/official-xeon-x5660-x58-review-discus...
"Overclocking the pcie bus makes a lot more sense now, my 950 pro went up to 1920/977 from 1670/920."

October 18, 2016 | 11:55 AM - Posted by tbone8ty (not verified)

how is the average home user/gamer going to benefit from these?

seems like its out of my class in terms of usage case.

would I see a noticeable upgrade from a 850 evo sata drive?

its still nice to see tech improve like this though. its the new gold standard.

October 18, 2016 | 12:01 PM - Posted by Allyn Malventano

My intention was to include some SATA products in the comparisons, but the 960 PRO was so fast that it explioted a bug in my code that required recompiling my tools and starting testing over, which was a shame because I had 60+ devices tested over the past several weeks. The four products in this review are the first four to complete the revised suite. More will follow shortly in additional reviews. I have a backlog of reviews to crank out as I can get additional samples tested.

If you do anything above light browsing, you'd probably notice a difference over an 850 EVO, but I would recommend holding off for the 960 EVO, which is supposed to be similarly competitive on price but should offer greater performance than the 950 PRO.

October 19, 2016 | 03:47 PM - Posted by Anonymous (not verified)

Allyn, you're a fucking Badass.

October 18, 2016 | 12:11 PM - Posted by Anonymostly (not verified)

I was wondering this myself.

Probably any ssd (sata or nvme) will improve anyone's computing experience over an hdd, but at what point do "regular" users no longer need to bother with faster SSDs?

And how far up the prosumer ladder do you have to climb in order for nvme drives (and particularly a beast like the 960 pro) to have an impact on one's experience?

For instance, I wish I could make Photoshop and SourceTree (git) faster and snappier, and task manager does show them as taxing the disk to a considerable extent, but at what point am I just wasting money unless I spend my days extracting 50GB archives for the sake of testing my new SSD's performance?

October 18, 2016 | 12:19 PM - Posted by Allyn Malventano

If you see your disk as more of the limit (vs. CPU thread saturation) for a given workload, then you should see an improvement with the faster SSD. Moving from any SATA to a 960 PRO is likely going to double to triple all of the storage metrics. It will probably shift the bottleneck back onto the CPU for anything that was previously storage limited.

November 19, 2016 | 03:32 PM - Posted by DarkUltra (not verified)

You could look at Disc Active time in the task manager. If it constantly or occasionally goes to 100% it means the CPU is waiting for the drive. On my PC with an 850 EVO sata SSD I rarely see this.

October 18, 2016 | 12:00 PM - Posted by encryption (not verified)

Should we expect a later update regarding encryption (OPAL 2.0 / IEEE 1667) that was promised for the 960 review in the comments for the 960's announcement?

October 18, 2016 | 12:03 PM - Posted by Allyn Malventano

We ran out of time to evaluate it as our sample arrived on Friday and we still do not have updated drivers or Magician software, but here is what was provided to us:

Advanced data encryption
The 960 PRO provides the same data encryption features as other Samsung SATA SSDs. Self-Encrypting Drive (SED) security technology will help keep data safe at all times. It includes an AES 256-bit hardware-based encryption engine to ensure that your personal files remain secure. Being hardware-based, the encryption engine secures your data without performance degradation that you may experience with a software-based encryption. Also, the 960PRO is compliant with advanced security management solutions (TCG Opal).
* Supporting eDrive(IEEE1667) is under consideration

October 18, 2016 | 12:36 PM - Posted by encryption (not verified)

Completely understandable, I was just wanting to make sure that we would eventually see follow-up on the topic. I know that a number of people, myself included, are upset that we're >1 year in and still don't have eDrive support on the 950 Pro that was originally promised, so I remain hopeful they will be taken to task on the topic. :)

October 18, 2016 | 12:38 PM - Posted by Chaitanya Shukla

Did you guys run thermal tests? Just curious to know how much that copper label works in helping dissipating heat from SSD?

October 18, 2016 | 12:43 PM - Posted by Allyn Malventano

It's on my list. Ryan is out of town and I couldn't locate our FLIR. Also, the testbed was crunching away on comparison samples and I didn't want to interrupt / restart a run. I will say that normal use on a 950 PRO doesn't heat it up enough to throttle, so it's safe to say there is even less concern on the 960 PRO.

October 18, 2016 | 03:16 PM - Posted by Jann5s

When you get to it, and if you have the time, and if you think it is interesting, I would be really interested in the way the drive behaves in a confined environment like on the back of an ITX motherboard, with nearly zero airflow.

Just to clarify, the case I have in mind is the A4-SFX, which has the GPU on the other side of the thin metal sheet. I'm worried that temperature at that location will be high enough (from the CPU and GPU heat) that the drive will always throttle. The throttling method can itself be interesting, if it is well managed and stable or very erratic.

Again, I understand this is a niche question, please ignore it if you think the audience is too small. In that case I will investigate myself once I get the case.

October 18, 2016 | 06:10 PM - Posted by Allyn Malventano

That's actually the conditions you'd need to risk overheat on this one. Normal use doesn't get these things much higher than warm, as they are idle most of the time. The faster performing ones actually heat up *less* with the same intermittent workload, since they complete the requests faster and return to idle state quicker.

October 19, 2016 | 11:52 AM - Posted by Chaitanya Shukla

Thanks will keep an eye out for update. I have a 950Pro in my PC which I mostly use for photo editing(raw editing, and macro stacking) and development work with VMs. In last year of ownership of that SSD only time I had issues with thermal throttling were during summers here in west India when temps crossed 45 Deg C(Antec Spot cool took care of that). I remember Samsung pointing out to that copper labels in press release of 960 Pro so just was curious how effective that solution has been for Samsung.

October 18, 2016 | 12:50 PM - Posted by Paul A. Mitchell (not verified)

FYI: Syba make this 2.5" U.2 (SFF-8639) to M.2 NVMe enclosure:

http://www.sybausa.com/index.php?route=product/product&product_id=884&se...

I don't see any thermal transfer material in the parts list.

If the enclosure does not contact the top side of this M.2 SSD, perhaps a custom thermal transfer pad could be added without doing any harm?

The 2.5" storage "ecosystem" is ubiquitous now, so this Syba enclosure looks like a really good way to cool these high-performance M.2 SSDs.

Of course, this solution also requires compatible cabling, which will increase the cost; but, the added cooling may be worth it.

In my mind, I'm building a RAID-0 array with four of these Syba enclosures, four compatible cables, and the RocketRAID 3840A NVMe RAID controller. This is ITEM #1 in my wish list to Santa Claus next December :)

October 19, 2016 | 12:30 PM - Posted by Allyn Malventano

It doesn't matter that much really. Plenty of heat transfers through the PCB and to the bottom of the SSD. Maybe even more than to the top since you'll get thermal conductivity to the pads, etc. 

October 18, 2016 | 12:41 PM - Posted by Paul A. Mitchell (not verified)

Allyn, As always, many thanks for sharing your extraordinary expertise with the rest of us. You be THE BEST, man!

October 18, 2016 | 01:16 PM - Posted by Paul A. Mitchell (not verified)

If you're a "number junkie" (like me),
here are some parametrics to chew on:

MAX HEADROOM of a single NVMe M.2 port is:

32 Gbps / 8.125 bits per byte = 3,938.46 MBps
(PCIe 3.0 "jumbo frame" is 130 bits / 16 bytes)

ATTO reports 3,460 MBps with this 2TB 960 Pro:

3,460 / 3,938.46 = 0.8785 or 12.1% overhead (!!)

Assume another 10% aggregate overhead
from a RAID-0 array of four of these SSDs:

3460 x 4 x 0.90 = 12,456 MB/second

PRETTY AMAZING! 12.4 Gigabytes per second (!!)

Even if the aggregate controller overhead is more,
approaching a raw READ speed of 10 GB per second
is right in the same ballpark as DDR3-1600 DRAM:

1600 x 8 = 12,800 MB/second (raw bandwidth)

Conclusion: we have now reached a new era
in which mass storage is capable of performing
at close to the same sequential performance
as volatile DDR3 DRAM. Four such M.2 SSDs
in RAID-0 mode == ~8TB (before formatting).

"Driver, can this bus go any faster?" asked the church mouse.

October 18, 2016 | 01:21 PM - Posted by Paul A. Mitchell (not verified)

Source of ATTO measurement of 3,460:
http://www.legitreviews.com/wp-content/uploads/2016/10/960pro-atto.jpg

October 18, 2016 | 01:50 PM - Posted by Paul A. Mitchell (not verified)

These comparative graphs are truly amazing:

http://hothardware.com/ContentImages/Article/2541/content/att1.jpg
http://hothardware.com/ContentImages/Article/2541/content/att2.jpg

October 18, 2016 | 10:12 PM - Posted by Allyn Malventano

Be careful with those. Folks have a habit of running them on empty drives.

October 19, 2016 | 02:46 PM - Posted by Paul A. Mitchell (not verified)

copy that: thanks!

MRFS

October 18, 2016 | 01:53 PM - Posted by Paul A. Mitchell (not verified)

To elaborate our comparison with DDR DRAM:

DDR2-800 x 8 = 6,400 MB/second
DDR3-1333 x 8 = 10,664 MB/second
DDR3-1600 x 8 = 12,800 MB/second

October 18, 2016 | 10:00 PM - Posted by Qep (not verified)

Fail PCPer! 600p failed because it sucked, not because the tests were bad. Atleast keep the file copy and file creation tests!

October 18, 2016 | 10:57 PM - Posted by Allyn Malventano

File creation was limited by the legacy tool that created the files. File copy effectively made duplicates of the same files on the same drive - something that people generally don't do. It was time to move on. 600p still looks to be plenty good for the money.

October 22, 2016 | 05:55 PM - Posted by muffinman (not verified)

I'm assuming they never patched the 600p to make it work correctly. Just watched the PC Per Podcast 416.

October 19, 2016 | 01:17 AM - Posted by Dudley Storey (not verified)

Excellent review, Allyn - thanks so much. To me, the next obvious step is to test two of the 960s in a RAID configuration, as you did with the 950 Pros: any possibility of trying that?

October 19, 2016 | 02:50 AM - Posted by BT (not verified)

GIMME, GIMME, GIMME...When do these hit the shelves? Do you
think the $129 price on the 250gb evo will still be offered do to a high demand?

October 19, 2016 | 02:16 PM - Posted by UnderwaterSurfer (not verified)

You can go to the Samsung site and pre-order the 250GB 960 EVO for $129.99, and they list "Ships in 6 to 8 weeks".

http://www.samsung.com/us/computing/memory-storage/solid-state-drives/ss...

October 19, 2016 | 12:12 PM - Posted by Jann5s

Seeing you do all this in excel makes my eyes hurt, why not use something like python to process these massive chunks of data?

October 19, 2016 | 12:33 PM - Posted by Allyn Malventano

I'd considered python for the first level of parsing, but Excel surprised me and actually handled it quite well. 

October 19, 2016 | 10:12 AM - Posted by Anonymous (not verified)

"•Hexadecimal Die Packages" really this is a bit overboard with the technobabble for technobabble's sake! Why not just say 16. Really 0x000F die packages man those marketing monkeys! What's next everything in octal that would be 17 in octal!

October 19, 2016 | 12:32 PM - Posted by Allyn Malventano

It probably just stems from how Samsung labels them on their part number decoder. 

October 19, 2016 | 02:48 PM - Posted by Paul A. Mitchell (not verified)

00011101010010101010
AND DON'T FORGET ...
00002220101010101111
GOT IT?? :)

October 19, 2016 | 01:13 PM - Posted by Joholo (not verified)

Great article, Allyn! Got a few questions about the new testing suite. In the article you said that current testing tools are not well suited for hybrid SSD's. How do you create a new suite? Do you program it yourself or do you use different tools?

I'm curious about it because the SSD reviews at PC Perspective are by far the best and most in depth. Most of the other review sites use "standard" tools like ATTO Disk Benchmark, AS SSD etc.

October 26, 2016 | 07:35 PM - Posted by Allyn Malventano

The code is in-house. It started with the work I did to get Latency Percentile results (which no benchmark does), and evolved into the need to apply custom workloads that better relate to how caching SSDs behave in real-world use. The workloads were designed from taking captures, but they are not playbacks of those captures. Think of it as synthetic trace, tuned to complete in a reasonable amount of time but with more realistic idle periods than an idle-cropped trace playback.

ATTO and ASSSD only test a small in-place file, with no adjustment for how allocated (full) the SSD is. Most people just run them on a fresh out of the box empty drive, and after looking at the results of this new suite, I've come to learn that empty SSDs tend to go much faster than even partially full ones.

October 19, 2016 | 06:17 PM - Posted by Thedarklord

How would you say the new 960 Pro 1TB M.2 compares to the 950 Pro 1TB SATA?

October 20, 2016 | 04:52 AM - Posted by Jann5s

I don't think the 950 Pro 1TB SATA exists. Do you mean 850 pro SATA? In both cases, the differences are enormous! :D

October 20, 2016 | 03:16 PM - Posted by Thedarklord

Edit: 850 Pro 1TB SATA
:)

October 26, 2016 | 07:36 PM - Posted by Allyn Malventano

The 850 PRO is similar to the EVO at 1TB.

October 20, 2016 | 05:23 PM - Posted by maonayze

I recently picked up the PM961 M.2 Nvme for £99 here in the UK. I believe this is the OEM version of the 960 EVO. Can you tell me if there are any differences between the two and could I use the driver for the 960 EVO and would it make a difference at all.

Love the site and the weekly podcasts, keep up the good work.

October 26, 2016 | 07:38 PM - Posted by Allyn Malventano

The 950 PRO driver is really just an optimized NVMe driver. The one that comes with Windows doesn't handle higher performing SSDs as well as it could. I haven't tried that driver on the PM961, but it should in theory work and might give it a bit of a bump in random write speeds.

October 23, 2016 | 12:39 AM - Posted by HERETIC (not verified)

Hi Allyn,
Nice to see where your heading with the test suite.
First things I look at with a SATA boot drive are-
1.Min of 8K iops read at QD1
2.Must be able to sustain 200MB write after the buffer.
(can't have it bottleneck my spinning rust)

Becoming difficult to find in the midrange now with small drives (128-256) the TLC race to the bottom-suspect we're
going to see worse once Ram-less drives hit-thro by next year
we might see the end of planer TLC...........................

November 16, 2016 | 09:08 PM - Posted by sensacion7

Wow. what a drive. I can't wait to test one soon !

November 19, 2016 | 03:55 PM - Posted by DarkUltra

So this drive has capacitors to help write out anything in flight if the power goes out:

https://youtu.be/nwCzcFvmbX0 skip to 2:00

Will the 960 Evo have that? Would it prevent something like this:

https://youtu.be/-Qddrz1o9AQ

March 6, 2017 | 08:49 AM - Posted by peter j connell (not verified)

Perhaps fanciful, but i agree it could be a killer app?

"Conclusion: we have now reached a new era
in which mass storage is capable of performing
at close to the same sequential performance
as volatile DDR3 DRAM. Four such M.2 SSDs
in RAID-0 mode == ~8TB (before formatting)."

My take on it would be a less ambitious 2 drive raid 0 of 512gm 960 ssds.Best performing and cheaper.

PCIe Gen 3.0 allows 1GB ps per lane, bidirectionally, so 2GB per lane theoretical max.

OR, 8GB ps for the 4 lane dual M.2 ports on moboS.

In theory thats sufficient to max out 2 raid 0 960 ssdS, but 3500MB ps sequential reads (writes are 2100MB), are of course unidirectional.

so in theory it seems raid 0 pair of 960s yields 4000MB sustained, read or write.

I am pretty sure we will see 8 lanes available to m.2 mobo sockets (even w/ bargain AMD Ryzen mobos & cpus (32 lanes BTW)), allowing 7000/4400 MB ps read write in theory, w/o fancy controllers.

I dunno the numbers for ram bandwidth. a lot better am sure. not sure thats a deal breaker for my argument.

point is, 7000/4400MB are numbers in a league of their own compared to anything before - even in the server world. Its a new paradigm for coders.

ok, using it for virtual memory isnt as fast as real memory, but shit its big. I dunno enough about architecture etc., but a TB of ram may open many possibilities for completely new approaches to old coding problems.

the killer benefit of ssdS was fast random access. It transformed our PCs.

~150MB ps sequential was livable, access times were the killer on HDDs performance.

As many have said re the 960, more of the same will be barely noticed by many.

give a gamer 1 TB of passable virtual memory, and apps which use it, then that could be revolutionary.

it bears repeating btw, that IOPS has shown even more stellar performance gains in the 960, and I imagine thats important for virtual memory. As we hear, many consider this the main reason to spend the extra for the 960 over the 950.

March 6, 2017 | 09:22 AM - Posted by peter j connell (not verified)

PS, upon reflection, poor mans raid 0 on 4 lanes is still attractive for swap/page files, even with little read speed gain. Write speed almost doubles from a theoretical 2200 MB ps to 4000MB ps.

Post new comment

The content of this field is kept private and will not be shown publicly.
  • Lines and paragraphs break automatically.
  • Allowed HTML tags: <a> <em> <strong> <cite> <code> <ul> <ol> <li> <dl> <dt> <dd> <blockquote><p><br>
  • Web page addresses and e-mail addresses turn into links automatically.

More information about formatting options

By submitting this form, you accept the Mollom privacy policy.