I'm looking at one of these to go along with my new build but I'm not sure which one to go for as I can't find a comparison review. I know the 950 is the retail version with a better warranty but on paper the SM951 seems to be higher spec but its a lot cheaper. Which is the one to go for?
Very interested to see if there are any real comparisons out there. Tempted to pick a 950 Pro up tomorrow if I come across any on sale.
The 950 is using 3D VNAND which consumes far less power than the MLC NAND in the 951. That's the major difference in them I believe. PC Per did a decent review and video.
What about performance though? on paper the SM951 is much faster than the 950 but is that actually the case?
Matching capacities (at least the 256GB varieties) trade blows depending on the workload and benchmark.
That's just annoying, so essentially it costs almost £40 for a retail box, two years extra warranty and lower power consumption. I find that a little hard to justify if that's all it is.
if you're looking at the 512GB models, the 950 also only has nand on one side of the card - d.t. the difference in nand type & die capacity... ...so the 512GB 951 is far more prone to suffer from thermal issues - which throttle speeds - unless you add active cooling. This is believed to be why there's only a 512GB model of the 950 thus far - until Samsung have production of 256Gb nand modules ramped up & can release a single sided 1TB version. Whilst m.2 cards have their uses for (compatible) portable devices, for people that actually have a usage that can materially benefit from the speeds, a u.2 version would be much more interesting imho - as using the case of a 2.5" drive as a heatsink (which is what most 2.5" & 3.5" SSDs do) should remove this issue entirely - plus allowing far higher capacities; potentially using cheaper nand... ...& throwing in a m.2-to-u.2 adaptor would add very little to the price.
AFAIR, the issue with the SM951 hitting thermal limits was more to do with there being no available documentation for setting the drive controller flags (it's on OEM device, so that would be dealt with by OEMs interacting with Samsung directly) meaning the drive controller kept trying to operate in FULL SPEED FULL POWER mode continuously, with nothing to tell it otherwise. The 950 Pro drive controller uses power states in the normal way you'd expect for any other consumer drive. You can make it throttle like the SM951 at extreme workloads, but unless your desktop happens to be running a live high-access database you're probably not going to be able to saturate it with a continuous high queue-depth workload outside of benchmarks. If you are worried about temperature, most m.2 slots are on the rear of the motherboard, so a simple thermal pad between the drive controller and the motherboard backplate provides MORE than enough heatsinking for the tiny 6.5W-7W peak loads of the SM951 and 950 Pro.
Took the plunge and got myself the 950 Pro today from Dabs as they just got it in stock today and cheaper than elsewhere. Should have it next week.
That's not contradicting what i said. What i actually wrote was having nand on both sides of the card increases the potential for thermal issues -> increased throttling... ...which is something else entirely from saying that the nand is the key component creating the heat. (this is slightly bitty as i'm dead on my feet atm) i'm not sure where you've seen that SM951 runs at full speed constantly; though it does run hotter than the 950 - see, for example - https://www.youtube.com/watch?v=d3GlInzvHr8 Now, on one hand, Samsung drives have had various versions of thermal throttling for years in all formats - so for example, the 2.5" & mSATA 840 drives have "Samsung’s Dynamic Thermal Guard Algorithm"... ...as do the 850 drives... ...though naturally the 2.5" ones were far less prone to it kicking in than with the mSATA ones. Whilst, on the other, both the 951 & 950 drives support the L1.2 pcie mode - so 'if' there's some issue with some boards using that to reduce idle power still lower that i can't immediately see, this logically should apply to both drives as they're using the same pcie power standard. Now, as to triggering thermal throttling, along with random r/ws, you can also trigger it on either drive with sequential r/ws (cooling dependant)... ...so, for example, you could reasonably have a batch process that - reads -> processes -> writes -> deletes original - that runs for long enough to do so; as well as some random r/w load. Either way though, this then comes back to why a consumer would want to pay a significant premium for either drive atm - since, under the argument that a low level usage won't trigger it then where exactly is the material benefit that would justify paying the extra anyway...? (whilst a heavy use that would justify it is, by the very nature of needing to run long enough to materially benefit, far more likely to trigger it) Or, to phrase it as a question, what exactly are these myriad of uses that all & sundry could have that would both materially gain from the speeds to justify the premium AND would never run for more than a matter of minutes (cooling dependant, this can be 2 minutes or less) so that there was never any thermal throttling???
Maybe I should clarify. From what I've read, the Nand itself doesn't get anywhere near as hot as the controller, so much so that it doesn't really contribute much to the overall heat.
Perhaps i need to re-explain - i've had some sleep so this should be reasonably coherent. (i hope) Now the issue isn't fundamentally that the controller gets too hot to operate - but that the heat from the controller spreads to the nand. So, if you've got a controller hitting +/-100C because there's insufficient cooling (either through lack of airflow &/or not having a case acting as a heat sink &/or...) then the nand can exceed it's operating temp - which can lead to both degradation of the nand &, d.t. the effect on charge leakage, potential data loss. This spreading of heat to the nand on the Samsung SSDs is shown in the YT clip linked to earlier (there's other videos about showing the exact same thing with other manufacturers' m.2 SSDs) - & is the key reason why there's thermal throttling; as the controller itself can keep on working at those kind of temps. Now, most ways of attaching m.2 (& msata) SSDs largely 'shelter'/'semi-cover'/whatever word you think is most appropriate some or all the rear side of the pcb where nand could be placed - so, whilst nand on the front can be kept slightly cooler by case airflow &/or simple radiating the heat away, the nand on the back simply won't be cooled to the same extent... Hence, having nand on the rear of a m.2 SSD makes it more prone to needing to thermal throttle to protect the nand under the same conditions.
By 'running at full speed', I mean the SM951's controller doesn't gracefully move up and down power states and clock speeds depending on temperature and usage, it will naively go from idle power then slam right up to the max power state and clock under any load, and then throttle back down once it hits the thermal limit. The UBX controller in the 950 Pro will try and move down to lower power states when under lighter loads in order to delay or prevent the controller heating to the thermal throttle temperature. The SM951 has this 'stupid' behaviour because it is an OEM drive, and the idea is the OEM will change those drive performance states to suit the thermal solution in the device they will be deployed in. Consumers don't have access to these tools, so the drive stays in it's default bare-bones power state setting.
Well they both use the same controller - as the UBX is the S4LN058A01... ...they both use L1.2 & so are reliant on a proper implementation of ASPM (on the motherboard side) in order to save the most power when idle; at the expense of a delay coming out of the L1.2 mode... [ASPM can be uni or bidirectional - & it's the latter that saves the most power & apparently 'can' simply not work stably on some motherboards.] ...& i can find nothing at all that suggests the f/w is dramatically different with regard to power states between the two. Now, it's obviously possible that i've missed something - so could you find a link that describes/demonstrates exactly what you're talking about, as i can find nothing at all relating to power states or 'modes' that's specific to the 951 & not the 950 (or vice versa)...? ...but accepting that the 951 runs slightly hotter, looking at the YT vid linked to earlier, both drives are following a very similar thermal pattern initially so there's no immediate evidence for what you're suggesting there. Otherwise, whilst i'm certainly not saying that power saving can't be important with regard to battery life with portable devices, the controller heating for very short periods isn't an issue of itself... ...as, again, it's the effect upon the nand that's the primary concern here & the heat doesn't transfer quickly enough to be an issue with low levels of i/o. The implication though would be that, if the controller's slowed down with low levels of i/o (>idle obviously) as you're describing, r/ws would take noticeably longer in that 'mode' with the 950 than the 951; & this doesn't appear to be borne out with any lighter tests... ...indeed it's somewhat the reverse (at least partly d.t. the nand difference); which again doesn't seem to support what you're saying. [by 'primary concern', i'm naturally not suggesting that the controller could run at any temperature - but that it's the effect on nand temps that can cause throttling under normal circumstances.]
NAND temperature isn't too much of a worry; the controller is the source of heat, and by the time it's managed to heat the NAND through PCB conduction to a temperature that would measurably affect the NAND the controller will long ago have failed due to tunnelling (and as we know the the JEDEC slide of SSD life vs. temperature, active temperatures correlate POSITIVELY with retention, not negatively). Besides, the thermistor that governs temperature throttling is on the controler die, not the NAND dies. As you can see in the video you posted, the difference in controller behaviour between the two drives is clear: the 950 Pro will slow itself down to maintain a constant steady-state level of performance. The stock SM951 will continue to bounce off the thermal throttling threshold by attempting to run at the maximum clock and voltage, throttling back when it hits the thermal limit, then jumping up once it has cooled to just below the throttling threshold value. In short-duration tests with idling in between the SM951 and 950 Pro will behave very similarly, the difference is noticeable during longer periods of load. The key issue is that the SM951 will only behave this way out of the box: OEMs will not be using it in this state (unless they're screwing up, or they have a dedicated thermal solution) and will instead be installing a customised firmware variant to the drive that will cause it to behave in a more '950 Pro-like' manner, dependant on the power and thermal budget available in the system it is intended to be used in. The 950 Pro, however, is a consumer drive. It has to assume a worst-case installation scenario (sitting in an unventilated case with no contact to any heatsinking material) so needs a firmware that can deal with that usecase by default.
Splitting your post & the response into two parts - as hopefully it'll make the response clearer... 1. First off, the starting point for my original comment was Anandtech - "The back side of the 950 Pro is empty so a 1TB model should be geometrically possible if not economical, but the extra NAND packages would be even more susceptible to thermal problems. Samsung is instead choosing to wait for their 256Gb third-generation V-NAND before offering a larger model of the 950 Pro." (http://www.anandtech.com/show/9702/samsung-950-pro-ssd-review-256gb-512gb/12) - so it's not something that i've randomly made up. 2. With that JEDEC table - (a) other than the two highlighted figures in green, nothing within it has ever been explicitly part of the JEDEC standards (b) it simply does not describe everything that can happen (c) &, most importantly, it is only an example model from intel. (my recollection is that specific data set is from ~2007/8 & was largely extrapolated, rather than representing real test data) So, whilst some of the overall conclusions may be worked to, it simply isn't correct to assume that the nand in every SSD will follow that exact pattern with those temperature ranges (see 4 below). Well, it would clearly be nonsense to suggest that increasing temps indefinitely would continue to yield improvements... ...& without going too far into extremes, with standard nand then heating to 250C for several hours is what's needed to refresh all of the cells to an almost new state by releasing all of the trapped electrons. The table also is only looking at the SSD running at specific constant temps - whereas, it's clear from the YT video that the nand can dramatically increase in temp in a relatively short time frame (again see 4 below). 3. Now, where increasing temp does help is with programing & erasing cells - as it takes less energy to do so & reduces oxide breakdown in these specific processes... ...however, most importantly, excessive heat (esp going beyond the specific nand rating) can shift the voltage thresholds within a nand cell out of their boundaries - this being the flip side of it being easier to program & erase cells. So, it's about striking a balance here - since too low a temp can reduce lifespan when programming or erasing by breaking down the oxide... ...but too high a temp can cause leakage of charge sufficient to alter the bit(s) stored in the cells. 4. Whilst nand varies (& i cannot immediately find a spec for the nand in either the 950 or 951) consumer nand is typically rated somewhere around a 0 to 65/70C range in operation; with an allowable change of ~20C per hour... ...whereas the UBX controller clearly both safely works to a far higher temp &, as the YT video showed, is both pushing the directly observable nand on the 951 (as you can't see the temps for the nand on the back of the 951) to temps to around or over the likely max & increasing by >>20C over a matter of minutes. 5. As to the thermistor not being on the nand (though there conceivably could be one or more elsewhere on the pcb than the controller), so what? When you're designing anything, you look at the effect that each component will have on the others - so having it on the controller doesn't matter one iota... ...& it certainly doesn't prove that the nand is the less temp sensitive component. 1. This isn't what you'd stated earlier that i was asking you to link to so that i could see what's going on - for example - "By 'running at full speed', I mean the SM951's controller doesn't gracefully move up and down power states and clock speeds depending on temperature and usage, it will naively go from idle power then slam right up to the max power state and clock under any load, and then throttle back down once it hits the thermal limit." - & you're now saying that, as opposed to 'any load' - "In short-duration tests with idling in between the SM951 and 950 Pro will behave very similarly, the difference is noticeable during longer periods of load." So which is it? 2. Looking at the YT video, both drives initially attempt to run at full speeds (which you can see from the transfer rates) - & both then throttle once they reach a certain temp... ...the difference between the two *in that YT video* is that, whilst the 950 will hit ~97C & then continuously throttle, the 951 will reach its max, then throttle, & once the temp drops sufficiently the speed will increase. Now, if you were to test the 950 within a machine with cooling such that, once it started throttling, the temps decreased to drop it to below it's temp threshhold - you would see exactly the same behaviour. So the 950 simply isn't doing anything magical with 'power states & clock speeds' - it's just got a lower temp at which it will throttle... ...& this is exactly the same behaviour you'll see with any SSD with thermal management; see, for example, Crucial's description of the M500 - "Adaptive thermal monitoring reduces NAND operations by approximately 40% when the SSD's internal measured temperature reaches 75°C. The SSD remains in this lowered performance state until the temperature returns to 65°C; the SSD then returns to normal operating mode. These transition points might occur at points somewhat different from case or ambient temperature because of thermal effects that cause a difference between temperature sensed at the PCB level and temperature at the specified operating environment." (https://www.micron.com/~/media/documents/.../m500_2_5_ssd.pdf) 3. Where is the evidence that OEMs are using custom f/ws or cooling with the 951? Anandtech's testing of the 951 (http://www.anandtech.com/show/9396/samsung-sm951-nvme-256gb-pcie-ssd-review) was on one found within an intel NUC... ...whilst the images of the NUC itself don't exactly show that there's the world of cooling - http://anandtech.com/Gallery/Album/4347#1 - so that kind of shows that big companies aren't all using custom f/w or cooling. That's not suggesting that you'd expect an OEM to have zero airflow (or not consider a heat sink with portable devices) if using a drive that gets exceptionally hot & throttles - simply that thermal throttling is part of how things are done now.
That's suspicion from Anandtech. More likely, Samsung want to wait for higher density vNAND to reduce production costs for a 1TB drive (not only do more modules = more cost, but routing the NAND interface traces through an already VERY dense compact PCB is no easy task), as 1TB m.2 drives are still a niche of a niche. The other issue is compatibility: m.2 double-sided boards have different clearance specifications than single-sided m.2 boards. While the individual values are not applicable to set-ups other than that tested, the trend in values certainly is, and is borne out by the physics of NAND storage. Those values shift naturally during operation of the NAND itself (because NAND is not held in a temperature controlled environment) The drive controller is responsible to compensating for voltage shift that occurs naturally in NAND storage, and indeed must in order for the drive to function correctly. Screwing up this voltage compensation behaviour leads to issues, such as the performance degradation suffered by the 840 EVO (where the voltage shift in older cells caused the compensation algorithm to throw a wobbler and take longer than necessary). It's not the opposite at all. Let me explain the behaviour again: The SM951 at default will attempt to maintain a maximum level of performance possible, with the controller operating at its highest performance state and voltage. The only things that will cause it to drop from this state is drive idle, or thermal throttling. This can be seen in the video; the drive will 'bounce' between the maximum performance state and the throttled state due to thermal hysteresis. The 950 Pro will also hit full performance, but the difference is when it hits that thermal threshold after time spent at full performance. It will gracefully drop down to a steady-state performance that is lower than the SM951, and can be sustained without throttling. Nope, you can see in the temperature/time graph that then SM951 is fluctuating between the max temperature and the throttle temperature. the 950 Pro drops back to a lower state and stays there rock-steady with almost no variance. Under NDA. If you're not ordering large numbers of drives from Smasung B2B, you won't get the full drive documentation needed for tuning. And no OEM that wants to maintain relationships with suppliers will provide it either. Actually, you can see the massive thermal pad coupling the SSD to the entire chassis to use as a heatsink quite clearly: You're forgetting that the primary market for m.2 drives is still laptops (and hybrid tablets), and that thermal limits there are only a byproduct of the REAL issue: power consumption. For a NUC plugged into the wall the drive can consumer as much [power as it wants (with a sufficient thermal solution), but in a laptop on battery you want that drive to be a LOT more conservative in its power states. But different OEMs with different devices will want that drive to perform differently: a slim ultraportable will have much more conservative power limits than a gaming-oriented 'desktop replacement'. The tricky part comes in telling the drive what performance state settings it should be using at a given time. You don't want to neuter performance when the laptop is plugged in. How this is achieved varies from OEM to OEM, and is a nonstandardised function. Pull a SM951 from a laptop that is doing some clever power management, and in the absence of that proprietary communication the drive will default to its max-performance-all-the-time setting. This is why you can have a drive that will test in one way on a laptop that is connected to a charger, another on battery, and possibly yet another (of that laptop had some thermal limitation) when placed in a different system.
Quickly typing to look at main topics as opposed to trying to cover every last detail. Part 1 - nand temps. What you're doing is making an assumption that at >55C, the retention will continue upwards indefinitely - which both isn't shown in the table & simply isn't true... ...since, without stating that it's literally a bell curve (as there's no evidence for this being the case or not), very high temps to reset the oxide to a nearly new state with the resulting effect on data means that there will be a nominally bell like curve. This curve will then be shifted left or right by the type of nand - ie the structure (as v-nand's different from standard nand), the cell size (inc oxide thickness), the number of bits stored, etc... ...& obviously consumer nand is rated to a narrower temp range than industrial nand is. The active test then assumes that a consumer drive is only on for 8 hours a day at 40C - which will shift the curve again... ...& so on... We also know that whilst high temps can anneal the oxide, this process can also cause voltage shifts sufficient to cause bit flipping... ...& also that high temps can cause oxide rupture (& there's no data as to whether the effect of lower temps increasing oxide stress in p/e cycles or high temps with oxide rupture is the more detrimental). Now, beyond all of those effects & others, whilst i can't find a rating for the 951, the 950 is only rated for operation between 0-70C... (http://www.samsung.com/global/busin...SSD/global/html/ssd950pro/specifications.html) ...however the controller has both max & throttled temps higher than that - which would be unusual if it was the temp sensitive component that's limiting the rating. So what would you suggest is the component that is the limiting factor then? &, if it's the nand (there's not much else it can be off the top of my head - as it's clearly not the solder or the pcb as the controller clearly is allowed to get hotter), couldn't it be the case that having nand on both sides makes it more likely that the nand will exceed its rate spec? Part 2 - the 951 vs the 950. Okay, so let's assume that you have access to some info that i don't about OEM options - but without any evidence it's supposition that there's magical extra modes or f/w... ...though, looking a the specific example, given that the fan in that NUC is underneath the mobo, i'm not entirely convinced that a thermal pad is going to provide fabulous cooling (well, the number of fanless coolers with thick thermal pads for CPUs that can hit >110C is vast). As to the claim about the 950's controller throttling & staying at that level irrespective of any temp fall, what's your explanation for this graph; where a fan is on, turned off & then turned on again - (http://www.legitreviews.com/samsung-ssd-950-pro-512gb-nvme-pcie-ssd-review_174096/3) Well, once throttled, the transfer rate is neither constant - fluctuating between 753 & 879MB/s... ...&, once the controller's sufficiently cooled again, it stops throttling.
