The power consumption is a lot higher than many drives today; especially idle power consumption. This is saying desktop or miniITX, but not laptop/tablet. It pulls more juice than the Intel Core Y processors - think about that.
That's surprisingly slow - plain dual-channel haswell (nothing cutting edge) ddr3 1600 should be getting around 23GB/s read, write or copy bandwidth. 8GB/s sounds like there's a bottleneck somewhere else (which, with those numbers, wouldn't really surprise me), or that memory is copied 3 times (which also wouldn't really surprise me) in the process of getting to the "disk".
And with a latency of less than 100 nanoseconds for ram vs. 1ms for an SSD, there's no chance of such SSD's replacing ram for anything vaguely performance critical.
What's the point of a RAM Disk anyways? Modern operating systems are pretty smart about caching data in RAM, the RAM Disk will just take away memory from that (Linux tmpfs is a bit smarter about it).
There are however NVMe PCIe cards with DRAM and Flash on them which actually offer durability.
Everything you could want from a SSD in 2015. Still, QD1 4k performance will ensure in day-to-day usage these won't perform much better than the current SATA drives. Pricing is also reasonable. I would have liked these to be priced similar to 850 Pro, pushing the 850 Pro price down, but the markup Samsung added is minimal. I might get one of these is supply won't be an issue.
Steady performance is on the look out. If it matches Intel on that then I might get one of these. Its great to see Samsung bringing an official consumer grade NVMe + V-NAND on an M.2 2280 form factor drive.
For home use or even workstations I'd rather care for burst performance than steady state. The reason is simple: at 1 - 2 GB/s the bursts are finishing VERY quickly. And the total amount of transfers does not magically increase as the drive becomes faster. I.e. it's going to be idle for more often than slower drives under medium load, which means it won't reach steady state as "quickly" as other drives. Unless one benchmarks strictly for that.
Getting a drive into steady state is much more related to write workload, and whether or not you are reserving some extra free space by making your partitions smaller. Not really related to the speed of the drive.
It is related to the speed if you're not simply pushing as many IOPS as possible for as long as possible, but rather have a real world task to accomplish. If the drive finishes those requests faster it will return to idle quicker and be able to run garbage collection, when a slower drive would still be hammered by the load. The garbage collection will also be faster, i.e. need less idle time.
Of course this assumes that one task does not contain more 4k writes than the drive has spare area and clean free space and that there will be a break before the next such request arrives. If this doesn't hold - either your software is very stupid for writing many GBs in 4k chunks or you very likely have a server database under heavy load. Which is not what I consider "home use or even workstations".
Why is power so high? A benefit of SSD over HDD is lower power use. But these power figures don't look compelling for laptop use.
In terms of the general death of 2.5" SATA drives, I really hope it doesn't happen any time soon. These new form factors don't provide enough to warrant displacing SATA SSD in probably 95% of use cases. For realistic laptop work loads, is SATA really a bottleneck? Really? I've not noticed any of my laptops pushing the limits of even SATA-2 in normal usage, with bottlenecks tending to be more the random access pattern of data, and for most users, this will be at a low queue depth and be lucky to push the envelope of even SATA-1.
I think these drives are tuned for Z170 desktops where drive power usage isn't a concern and where enthusiasts are just after maximum performance. I'd expect Samsung to announce a different M.2 SSD that will be aimed at laptop usage.
You are totally missing the point. NVMe opens for lower latency, wich can increase 4K QD1-5 performance... where most consumer and workstation loads are. The sequential read/writes isn't that interesting.
Power is up 40%ish (5 to 7W) but IOPS is up 3x. IOPS/w is much higher. You're going to have a race-to-sleep situation here, just like we have with CPUs.
Nah 2.5" sata SSD's will probably stay around for quite some time, especially in the value markets. I would expect most of them to be TLC before long, though.
For OS and app loading the difference should be very noticeable. For other activities it depends what your usage is. The nice thing about M.2 is that a regular sized laptop can accommodate both an M.2 drive for those places where it can excel and a 2.5" platter drive for sheer capacity.
I'm hoping Microsoft implements something like Apple's Fusion Drive function, which treats the combined capacity of an SSD and a platter drive as a single volume and decides behind the scenes which files should live where based on their type and frequency of access. The affordability of larger SSDs has reduced the need somewhat but images and video keep getting bigger. Not having to think about what should go where is something the OS should provide for the user. Too many of them are simply not equipped to give the issue proper consideration, so if it is automated it's better for everyone.
Some boards already offer this as a software solution I think - it's not an actual hardware implementation just a "value add" that is locked to the installed board.
Dissipating 5 - 7 W on such a small card is going to get toasty under sustained load. Anyone interested in that (and not using a laptop) could add small DRAM coolers, which should be enough. Under bursty real world loads temperature shouldn't be a problem, though.
Samsung says they've got new thermal management algorithms so we shouldn't see performance affected much, but I'll definitely have to keep an eye on that while testing.
Users won't be doing much sustained loading of these drives while they have 512GB capacities only. You'd literally fill the drive if you sustained a transfer for any length of time that would heat up the drive enough for it to throttle.
Yes, how badly these throttle under heavy sustained transfers will be the question. I don't think there is any way to fit even a very small heatsink on the controller. The M.2 form factor just doesn't leave any room for things like that.
Perhaps if the UBX controller is facing upwards when installed in a desktop M.2 slot, then you could do it. Laptop and SFF/miniITX owners with an M.2 trapdoor or M.2 underneath the board will have no luck.
This is great. We have finally reached the 2.5GB/s. I am hoping we top the PCI-Express 3.0 x4 soon up to 3.2GB/s. And the switch to NVMe.
I dont think any more performance improvement over this will be much of a selling point. We will likely reach the stage of SATA3 where pretty much all SSD on the market have similar performance in Seq and acceptable Random.
But i am concern about the power usage. Idle Power usage is way too high. I hope the next battle field will be here, where Laptop can squeeze little more time out of it.
How much of those juice are going to controller? And how many of those goes to NAND? They already switched to LPDDR4, which is more power efficient over LPDDR3 already. And what percentage of power goes to RAM?
As much as this Anandtech article mentioned that these will be for laptops, I see these more as a play to take advantage of all of those new Z170 desktop motherboards that recently hit the market. Pretty much all of them have M.2, PCIe 3.0 x4, NVMe capable slots to fill, and there is a lack of available drives to do it with.
On the desktop front these drives make sense. Maximum performance with power being an after thought. In a laptop or an ultrabook, the power draw is a significant issue.
Love these news. 2,500MB/s read and 1,500MB/s write using NVMe will be the mainstream soon. I don't think non-enterprise consumers need more throughput than that for the next decade. The goal now should be to lower GB/$ for this kind of storage. I'm not sure how long nand-based flash drive will last before next technology comes, but really, I'm happy with the progress made on flash drives.
That's a rather significant increase in idle power consumption. I'm not worried about average and "burst" power since the drive makes up for it in actual speed, but any idea why idle is so high in comparison to SATA? Is it the interface?
The interface probably has a lot to do with it. High-speed PHYs are power hungry. Putting the device to sleep saves a ton of power, at the cost of increased latency waking it up.
There's probably not much incentive for Samsung to work on improving the idle power usage. Laptops are always going to need to be aggressive about putting the drive to sleep, and desktop users aren't going to care about half a watt of difference.
Is there a consensus on what types of connectors we will be seeing on consumer/business enthusiast and workstation systems over the next 24 months? I have no idea when the next generation of storage controller will come out but is it reasonable to assume a doubling in speed to 5GB/s reads? That would saturate an M2 pcie connector, not such a problem if the M2 becomes pcie 4.0 x 4 (8GB/s transfer if my napkin wiki math is correct.) Would it make sense to have a mixture of two M2 pcie two NVMe and four to six SATA that would seem to me to be a good balance without depriving pcie lanes for other duties...
Compared to the 40MB IDE harddisk I had in my first PC the hardware we have access to now is insane :) How much internal bandwidth would for example a video editor be using on a high end workstation, a 5960 with 32+GB RAM a couple of 1.2TB Intel 750 drives as scratch drives rendering in Premier via one or two Geforce Titans and the resulting product being pushed out to a SAN via 10GbE, when will 40 lanes not be enough :)
Articles like this make me glad Im in this industry, there is always something coming down the pipeline :)
I think M.2 is the form factor of choice for high-performance consumer SSDs, for the Skylake generation platform at least. Using the U.2 connector to provide PCIe x4 to a 2.5" drive will probably be more popular than SATA Express was, but it's still going to be a very small niche from a consumer perspective. Any wider/faster drive interface is still a long ways out, because products needing that speed can just be put on a PCIe card.
I really wonder about the heat. All the m.2 drives run hot and nvme runs hotter (Intel said they didn't put out an m.2 version of the 750 SSDs because they found NVME made them run too hot). On a lot of mobos, these things lie almost flush to the pcb, so not much room for air circ. The heat sink on this thing doesn't look too impressive either.
Will the 950, like the 850, have 40 nm cells? Presumably yes, since Samsung is claiming comparable durability. 40 nm sounds much more trustworthy than 19 nm; data retention gets dicey when you can count the number of trapped electrons on the finger of one hand.
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.
49 Comments
Back to Article
HollyDOL - Tuesday, September 22, 2015 - link
Looks sexy... now let's see if price and benchmarked performance are appealing as well :-)eanazag - Tuesday, September 22, 2015 - link
The power consumption is a lot higher than many drives today; especially idle power consumption. This is saying desktop or miniITX, but not laptop/tablet. It pulls more juice than the Intel Core Y processors - think about that.SunnyNW - Tuesday, September 22, 2015 - link
How does m.2 x4 Nvme speeds compare to Ramdisks?Lolimaster - Tuesday, September 22, 2015 - link
Not even change, DDR3/DDR4 ramdisks delivering 8GB/s but more importantly 10x more iops, 100x less response time.emn13 - Friday, September 25, 2015 - link
That's surprisingly slow - plain dual-channel haswell (nothing cutting edge) ddr3 1600 should be getting around 23GB/s read, write or copy bandwidth. 8GB/s sounds like there's a bottleneck somewhere else (which, with those numbers, wouldn't really surprise me), or that memory is copied 3 times (which also wouldn't really surprise me) in the process of getting to the "disk".And with a latency of less than 100 nanoseconds for ram vs. 1ms for an SSD, there's no chance of such SSD's replacing ram for anything vaguely performance critical.
nils_ - Sunday, September 27, 2015 - link
What's the point of a RAM Disk anyways? Modern operating systems are pretty smart about caching data in RAM, the RAM Disk will just take away memory from that (Linux tmpfs is a bit smarter about it).There are however NVMe PCIe cards with DRAM and Flash on them which actually offer durability.
bug77 - Tuesday, September 22, 2015 - link
Everything you could want from a SSD in 2015. Still, QD1 4k performance will ensure in day-to-day usage these won't perform much better than the current SATA drives.Pricing is also reasonable. I would have liked these to be priced similar to 850 Pro, pushing the 850 Pro price down, but the markup Samsung added is minimal. I might get one of these is supply won't be an issue.
fallaha56 - Tuesday, September 22, 2015 - link
Except the 1TB capacity...Metroid - Tuesday, September 22, 2015 - link
Steady performance is on the look out. If it matches Intel on that then I might get one of these. Its great to see Samsung bringing an official consumer grade NVMe + V-NAND on an M.2 2280 form factor drive.MrSpadge - Tuesday, September 22, 2015 - link
For home use or even workstations I'd rather care for burst performance than steady state. The reason is simple: at 1 - 2 GB/s the bursts are finishing VERY quickly. And the total amount of transfers does not magically increase as the drive becomes faster. I.e. it's going to be idle for more often than slower drives under medium load, which means it won't reach steady state as "quickly" as other drives. Unless one benchmarks strictly for that.extide - Tuesday, September 22, 2015 - link
Getting a drive into steady state is much more related to write workload, and whether or not you are reserving some extra free space by making your partitions smaller. Not really related to the speed of the drive.MrSpadge - Tuesday, September 22, 2015 - link
It is related to the speed if you're not simply pushing as many IOPS as possible for as long as possible, but rather have a real world task to accomplish. If the drive finishes those requests faster it will return to idle quicker and be able to run garbage collection, when a slower drive would still be hammered by the load. The garbage collection will also be faster, i.e. need less idle time.Of course this assumes that one task does not contain more 4k writes than the drive has spare area and clean free space and that there will be a break before the next such request arrives. If this doesn't hold - either your software is very stupid for writing many GBs in 4k chunks or you very likely have a server database under heavy load. Which is not what I consider "home use or even workstations".
eddieobscurant - Tuesday, September 22, 2015 - link
Does it have an option rom like the kingston hyperx predator or is it like the sm951 which needs motherboard support in order to boot?I suppose the latter but it would be nice if it has an option rom for booting.
Ryan Smith - Tuesday, September 22, 2015 - link
The latter I believe.TheWrongChristian - Tuesday, September 22, 2015 - link
Why is power so high? A benefit of SSD over HDD is lower power use. But these power figures don't look compelling for laptop use.In terms of the general death of 2.5" SATA drives, I really hope it doesn't happen any time soon. These new form factors don't provide enough to warrant displacing SATA SSD in probably 95% of use cases. For realistic laptop work loads, is SATA really a bottleneck? Really? I've not noticed any of my laptops pushing the limits of even SATA-2 in normal usage, with bottlenecks tending to be more the random access pattern of data, and for most users, this will be at a low queue depth and be lucky to push the envelope of even SATA-1.
r3loaded - Tuesday, September 22, 2015 - link
I think these drives are tuned for Z170 desktops where drive power usage isn't a concern and where enthusiasts are just after maximum performance. I'd expect Samsung to announce a different M.2 SSD that will be aimed at laptop usage.JKJK - Tuesday, September 22, 2015 - link
You are totally missing the point.NVMe opens for lower latency, wich can increase 4K QD1-5 performance... where most consumer and workstation loads are.
The sequential read/writes isn't that interesting.
JKJK - Tuesday, September 22, 2015 - link
And I point out again... LATENCYBillyONeal - Tuesday, September 22, 2015 - link
Power is up 40%ish (5 to 7W) but IOPS is up 3x. IOPS/w is much higher. You're going to have a race-to-sleep situation here, just like we have with CPUs.neo_1221 - Tuesday, September 22, 2015 - link
True, but idle power is 4x higher than the 850 Pro - definitely a concern if you're thinking of putting this in a laptop.extide - Tuesday, September 22, 2015 - link
Nah 2.5" sata SSD's will probably stay around for quite some time, especially in the value markets. I would expect most of them to be TLC before long, though.epobirs - Sunday, September 27, 2015 - link
For OS and app loading the difference should be very noticeable. For other activities it depends what your usage is. The nice thing about M.2 is that a regular sized laptop can accommodate both an M.2 drive for those places where it can excel and a 2.5" platter drive for sheer capacity.I'm hoping Microsoft implements something like Apple's Fusion Drive function, which treats the combined capacity of an SSD and a platter drive as a single volume and decides behind the scenes which files should live where based on their type and frequency of access. The affordability of larger SSDs has reduced the need somewhat but images and video keep getting bigger. Not having to think about what should go where is something the OS should provide for the user. Too many of them are simply not equipped to give the issue proper consideration, so if it is automated it's better for everyone.
nils_ - Sunday, September 27, 2015 - link
Some boards already offer this as a software solution I think - it's not an actual hardware implementation just a "value add" that is locked to the installed board.MrSpadge - Tuesday, September 22, 2015 - link
Dissipating 5 - 7 W on such a small card is going to get toasty under sustained load. Anyone interested in that (and not using a laptop) could add small DRAM coolers, which should be enough. Under bursty real world loads temperature shouldn't be a problem, though.Billy Tallis - Tuesday, September 22, 2015 - link
Samsung says they've got new thermal management algorithms so we shouldn't see performance affected much, but I'll definitely have to keep an eye on that while testing.dave2150 - Tuesday, September 22, 2015 - link
Users won't be doing much sustained loading of these drives while they have 512GB capacities only. You'd literally fill the drive if you sustained a transfer for any length of time that would heat up the drive enough for it to throttle.khanov - Tuesday, September 22, 2015 - link
The SM951 512GB has been shown to throttle in under 2 minutes:http://www.anandtech.com/show/8979/samsung-sm951-5...
khanov - Tuesday, September 22, 2015 - link
Yes, how badly these throttle under heavy sustained transfers will be the question. I don't think there is any way to fit even a very small heatsink on the controller. The M.2 form factor just doesn't leave any room for things like that.Perhaps if the UBX controller is facing upwards when installed in a desktop M.2 slot, then you could do it. Laptop and SFF/miniITX owners with an M.2 trapdoor or M.2 underneath the board will have no luck.
Daniel Egger - Tuesday, September 22, 2015 - link
Eek, that idle power is horrible.Billy Tallis - Tuesday, September 22, 2015 - link
The 1.7W idle is when it's still fully on, but not processing commands. When put in device sleep mode, Samsung says it'll draw 2mW.iwod - Tuesday, September 22, 2015 - link
This is great. We have finally reached the 2.5GB/s. I am hoping we top the PCI-Express 3.0 x4 soon up to 3.2GB/s. And the switch to NVMe.I dont think any more performance improvement over this will be much of a selling point. We will likely reach the stage of SATA3 where pretty much all SSD on the market have similar performance in Seq and acceptable Random.
But i am concern about the power usage. Idle Power usage is way too high. I hope the next battle field will be here, where Laptop can squeeze little more time out of it.
How much of those juice are going to controller? And how many of those goes to NAND? They already switched to LPDDR4, which is more power efficient over LPDDR3 already. And what percentage of power goes to RAM?
James5mith - Tuesday, September 22, 2015 - link
As much as this Anandtech article mentioned that these will be for laptops, I see these more as a play to take advantage of all of those new Z170 desktop motherboards that recently hit the market. Pretty much all of them have M.2, PCIe 3.0 x4, NVMe capable slots to fill, and there is a lack of available drives to do it with.On the desktop front these drives make sense. Maximum performance with power being an after thought. In a laptop or an ultrabook, the power draw is a significant issue.
jimjamjamie - Tuesday, September 22, 2015 - link
Get in my desktop!!texasti89 - Tuesday, September 22, 2015 - link
Love these news. 2,500MB/s read and 1,500MB/s write using NVMe will be the mainstream soon. I don't think non-enterprise consumers need more throughput than that for the next decade. The goal now should be to lower GB/$ for this kind of storage. I'm not sure how long nand-based flash drive will last before next technology comes, but really, I'm happy with the progress made on flash drives.extide - Tuesday, September 22, 2015 - link
Personally, I want higher GB/$ ;)lilmoe - Tuesday, September 22, 2015 - link
That's a rather significant increase in idle power consumption. I'm not worried about average and "burst" power since the drive makes up for it in actual speed, but any idea why idle is so high in comparison to SATA? Is it the interface?Billy Tallis - Tuesday, September 22, 2015 - link
The interface probably has a lot to do with it. High-speed PHYs are power hungry. Putting the device to sleep saves a ton of power, at the cost of increased latency waking it up.There's probably not much incentive for Samsung to work on improving the idle power usage. Laptops are always going to need to be aggressive about putting the drive to sleep, and desktop users aren't going to care about half a watt of difference.
BobSwi - Tuesday, September 22, 2015 - link
That's quite a cut in warranty, here's to hoping for some decent 850 Pro 512GB deals.hyno111 - Tuesday, September 22, 2015 - link
Source spec said it is using LPDDR3 memory..?BennettJ - Tuesday, September 22, 2015 - link
Agreed. Looks to be LPDDR3 based on other material online. Is this really using LPDDR4?willis936 - Tuesday, September 22, 2015 - link
Will my next build have an Xpoint boot drive and v-nand storage drive?WatcherCK - Tuesday, September 22, 2015 - link
Is there a consensus on what types of connectors we will be seeing on consumer/business enthusiast and workstation systems over the next 24 months? I have no idea when the next generation of storage controller will come out but is it reasonable to assume a doubling in speed to 5GB/s reads? That would saturate an M2 pcie connector, not such a problem if the M2 becomes pcie 4.0 x 4 (8GB/s transfer if my napkin wiki math is correct.) Would it make sense to have a mixture of two M2 pcie two NVMe and four to six SATA that would seem to me to be a good balance without depriving pcie lanes for other duties...Compared to the 40MB IDE harddisk I had in my first PC the hardware we have access to now is insane :) How much internal bandwidth would for example a video editor be using on a high end workstation, a 5960 with 32+GB RAM a couple of 1.2TB Intel 750 drives as scratch drives rendering in Premier via one or two Geforce Titans and the resulting product being pushed out to a SAN via 10GbE, when will 40 lanes not be enough :)
Articles like this make me glad Im in this industry, there is always something coming down the pipeline :)
Billy Tallis - Wednesday, September 23, 2015 - link
I think M.2 is the form factor of choice for high-performance consumer SSDs, for the Skylake generation platform at least. Using the U.2 connector to provide PCIe x4 to a 2.5" drive will probably be more popular than SATA Express was, but it's still going to be a very small niche from a consumer perspective. Any wider/faster drive interface is still a long ways out, because products needing that speed can just be put on a PCIe card.rochlin - Tuesday, September 22, 2015 - link
I really wonder about the heat. All the m.2 drives run hot and nvme runs hotter (Intel said they didn't put out an m.2 version of the 750 SSDs because they found NVME made them run too hot).On a lot of mobos, these things lie almost flush to the pcb, so not much room for air circ. The heat sink on this thing doesn't look too impressive either.
AnnonymousCoward - Wednesday, September 23, 2015 - link
how is it taking samsung so long to replace the 850 EVO with PCIe/NVMe??DocSportello - Wednesday, September 23, 2015 - link
Will the 950, like the 850, have 40 nm cells? Presumably yes, since Samsung is claiming comparable durability. 40 nm sounds much more trustworthy than 19 nm; data retention gets dicey when you can count the number of trapped electrons on the finger of one hand.OFelix - Friday, October 2, 2015 - link
When will you able to publish a review of this beauty?Flying_toaster - Sunday, October 18, 2015 - link
From where did you got the 1.7 W for idle power? In Samsung spec sheet is 70mW listet: http://www.samsung.com/global/business/semiconduct...ewitte - Friday, October 23, 2015 - link
I love the confidence when I ask about NVME support on an ASROCK Z87. Notice the word "should" :DHi Eric,
It should be work with latest BIOS.
However, nVME drives required UEFI mode Windows installation.
...