You are looking at a 33% reduction in flash price at best vs TLC on the same process. Complexities of manufacturing on the new process may reduce this. Requirements for a more capable controller and firmware to implement the more powerful error correction work against this cost savings.
Lets use the current market as an estimator. To remove as many variables as possible, let's look at the an MLC and TLC drive with the same controller from the same manufacturer. The 960PRO and 960EVO are current drives that meet this criteria. 960PRO - $329.99 (512MB) / $629.99 (1024GB) 960EVO - $249.99 (500GB) / $479.99 (1000GB) https://www.anandtech.com/show/10833/the-samsung-9...
I'm going to ignore the small difference in size (this actually makes TLC look better anyways). While TLC nand is able to fit ~50% more data on the same area of silicon, the MLC drive (PRO) only costs 32% (512GB) / 31.25% (1024GB) more than the TLC (EVO) drive. Alternately stated, even though the MLC nand holds ~33% less data on the same area of silicon, the TLC drive (EVO) only costs 24.24% (500GB) / 23.81% (1000GB) less than the MLC (PRO) drive.
Given this information it is reasonable to assume that a QLC drive might cost ~18 less than its TLC counterpart. 33 / 24.24 = 25 / X -> X = 25 * 24.24 / 33
A 20% price reduction at the same size is not exactly the kind of price reduction that displaces spinning disks. It is the progression of feature size reduction (planar NAND) and stacking (3D NAND) that is making flash more competitive. Furthermore, not everyone agrees that the 20% price reduction is of sufficient value to compensate the reduction in voltage drift tolerances (how far can the stored voltage move before flipping a bit) and possible performance deficits. The effective endurance of an active QLC SSD may compare to last generation TLC SSDs, but TLC with the same correction algorithms will have much higher effective endurance. Performance of QLC drives is currently an unknown quantity. Though many believe they can't possibly be slower than spinning disks (and they are likely correct), the first jmicron based SSDs were significantly slower than spinning disks, so it is possible that performance won't meet all expectations either. There is also a question of unpowered data retention and the effects of storing infrequently accessed data (periodic rewrites?, data retention issues?).
More important than the flash type is a useful endurance rating, which also includes the effect of the controller. A drive with twice as many P/E cycles doesn't help you if the controller causes 3 times the write amplification. Or: QLC with 1000 P/E cycles is more durable than TLC on an older controller with e.g. 800 P/E cycles.
Seeing that 1000 P/E cycle number I'm torn between hoping it's only marginally slower than TLC to minimize the performance hit when it inevitably takes over value SSDs, or hoping that its performance is so much worse than TLC that it's not viable for a general purpose SSD.
Can we start calling multi-level cell NAND dual-level cell NAND yet? Calling it multi-level doesn't make any sense with triple and quad-level NAND around.
>Calling it multi-level doesn't make any sense with triple and quad-level NAND around.
None of the name make any sense, since "triple level" actually has 8 levels, and "quad level" actually has 16 levels. They should have just gone with "2/3/4 bit per cell".
Does the thought of these QLC chips bring on a shudder of fear in anyone else? Yes they are starting at enterprise and thats fine, but you know these are going to end up in consumer SSDs, the Micron chips always do. This is not going to be understood by 98% of the market. OEMs will fill their laptops with them, retailers will be selling them at competitive prices. We saw so much bad TLC flood the market early on and while i give it to them that its got a lot better, anyone who knows quite how bad QLC will be should be quite worried.
I guess in a flash filled future there are simply going to be many grades of it in existence, unfortunately we just too often see a race to the bottom....
These do not fill me with dread , as long as they can keep the write amplification down, a feat I think they are perfecting from knowledge of 3bit cell. . Chips are getting smaller, and cheaper so yes all flash desktop PC, for everyone, should be here in 2020
recall that old dodge: "if you step halfway to the wall, how many steps does it take to get to the wall?" we're within cite of Heisenberg NAND. not something I'm thrilled about.
@dromoxen: "Chips are getting smaller, and cheaper so yes all flash desktop PC, for everyone, should be here in 2020"
Do you really think a ~20% price reduction at the same capacity vs a TLC drive is going to be the tipping point here? I think better stacking, which helps TLC and even MLC as well, is playing a bigger role here.
Heck no, I'm excited and looking forward to QLC SATA/AHCI for bulk storage. I'm waiting on drives to come out to start my homelab builds. Hell, I want them to be DRAMless also, to make BoM cheaper. We've got plenty of NVMe/PCIe drives now for cache & performance including improved V-NAND TLC (which lasts MUCH longer than you'd probably give credit to, upcoming Z-NAND (which acts like SLC) and also Optane (which I think is GREAT; looking forward to x4 laned M.2s). I see zero issue with QLC in laptops and other products. I'm guessing you're the type of person to only use or want a SINGLE drive for an entire system? Ehhh. I guess I see issue there then.
Forgot to add that a QLC NAND drive will still be better at random read & write, and bursts, than any HDD, 3.5" or 2.5" that budget products would have. And that's a performance region where the vast majority of said systems are working in. High queue depth, sequential, sustained operations are rarer.
@CheapSushi: "Forgot to add that a QLC NAND drive will still be better at random read & write, and bursts, than any HDD, 3.5" or 2.5" that budget products would have. And that's a performance region where the vast majority of said systems are working in."
I believe these will be targeted at what HDDs are currently targeted at. What you state is true if these are used as a primary storage device on a system that would otherwise have a HDD filling that role. However, in a secondary bulk storage role where there is a much lower frequency of access and accesses are much more sequential in nature, the superior random access characteristics may go unnoticed.
My concern is certainly not for me or you, but for the millions of consumers out there who do only run a single disk system. I fully agree in a a perfect world more levels of nand for people in the know is better, ive just seen how the market deals with this before and thats where my concern is.
Not read anything on what impact the new pci-e 4.0 and upcoming 2020 pci-e 5.0 versions will have on storage performance? New sata specs to leverage as much? As well as Product lines in the works that take advantage of the new versions? ( been stuck at pci-e 2.0 buying into x58... Incremental performance iterations since then has soured my enthusiasm for upgrading past my i7-980x for which remains quite adequate for my workstation needs .vrt pulling in zbrush... But since Coffee-lake introduced 6 cores providing an obviously sexy hedt alternative that matches my current core count... The upgrade enthusiasm is back )
However... I am in no hurry to buy into a defunkt PCI-E version again and suffer storage compromises ( like I have under marvell's oppressive sata 3 controller on the x58 for the past 8 years! On the same dam year PCI-E 3.0 released! Talk about bad timing )
Personally hoping the rumored 8-core Cannon-lake releases in April ( further blurring hedt/xeon skus ) with PCI-E 4.0 implemented in motherboard lines hopefully in addition to a slew of announcements from all the storage players touting new tech/product lines leveraging the bandwidth/thruput?
Sure seems making a major investment risky? After 8 years of 3.0... I would hate to be stuck waiting out another 8 years stuck at 4.0 because I jumped the gun in this clausterphobic release season right before 5.0 which again supposedly doubles bandwidth?
( would be like being stuck at 3.0 while tje rest of the world got to realize a comparably NVLINK advantage )
Ghetto resentment!
Trying to not get to excited for anything and just make tmp builds till 5.0...
Just seems to important of a timing concern to make major investments without consideration of 5.0 128GBps lanes compared to today's 32GBps or even stuck at 4.0's 64GBps bandwidth waxing for that 2x leap.
PCIe 4.0 is not even implemented yet and you're already talking about 5.0? 0_o I think you are putting too much faith in the industry if you think it will move that fast. I understand you do not want to "miss out" on the latest tech, but I feel you are misguided. If you look back at PCIe 2.0 to 3.0, very few applications actually use it. Go check benchmarks of recent graphic cards tested on Sandy Bridge and latest processors to see that your loss is more theoretical than real. Storage has moved from SATA3 to PCIe, but even then, the real usage difference between the best SATA3 SSD (which you can still use on your platform) and a more recent PCIe 3.0 SSD is negligible (unless you're running a storage heavy server load).
Conclusion: Upgrade when you need to, to an equipment you'll be happy with, not when the technology changes (unless you know a new product is literally right around the corner in a matter of weeks). The march of technology never stops and there is *always* something better coming (intel and nvidia have new products every year after all).
whether the P/E value is credible will be determined by the node used. current TLC on 3D is said to be on 40-50 nm, thus gaining charge count/cell. one might hope that QLC is at least on these nodes.
There was no plan to shrink the node for 3D NAND last I heard. See 3D NAND as a tower. They need a "base" of a certain size, otherwise, they risk it toppling over. So far, they just increase the number of floors to increase density (from 32 layers to 64 then 96).
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.
Facebook
Twitter
RSS
29 Comments
Back to Article
drajitshnew - Monday, February 12, 2018 - link
I just hope that they clearly differentiate between various flash types--even for enterprisesonicmerlin - Monday, February 12, 2018 - link
Bet they won’t.Still if these end up like half the price of normal SSDs then they’d be pretty useful for playing games and movies off of.
saratoga4 - Monday, February 12, 2018 - link
You're only gaining 1/3 more space, so there is no way you're going to get a 50% cost reduction.BurntMyBacon - Thursday, February 15, 2018 - link
You are looking at a 33% reduction in flash price at best vs TLC on the same process. Complexities of manufacturing on the new process may reduce this. Requirements for a more capable controller and firmware to implement the more powerful error correction work against this cost savings.Lets use the current market as an estimator. To remove as many variables as possible, let's look at the an MLC and TLC drive with the same controller from the same manufacturer. The 960PRO and 960EVO are current drives that meet this criteria.
960PRO - $329.99 (512MB) / $629.99 (1024GB)
960EVO - $249.99 (500GB) / $479.99 (1000GB)
https://www.anandtech.com/show/10833/the-samsung-9...
I'm going to ignore the small difference in size (this actually makes TLC look better anyways). While TLC nand is able to fit ~50% more data on the same area of silicon, the MLC drive (PRO) only costs 32% (512GB) / 31.25% (1024GB) more than the TLC (EVO) drive. Alternately stated, even though the MLC nand holds ~33% less data on the same area of silicon, the TLC drive (EVO) only costs 24.24% (500GB) / 23.81% (1000GB) less than the MLC (PRO) drive.
Given this information it is reasonable to assume that a QLC drive might cost ~18 less than its TLC counterpart.
33 / 24.24 = 25 / X -> X = 25 * 24.24 / 33
A 20% price reduction at the same size is not exactly the kind of price reduction that displaces spinning disks. It is the progression of feature size reduction (planar NAND) and stacking (3D NAND) that is making flash more competitive. Furthermore, not everyone agrees that the 20% price reduction is of sufficient value to compensate the reduction in voltage drift tolerances (how far can the stored voltage move before flipping a bit) and possible performance deficits. The effective endurance of an active QLC SSD may compare to last generation TLC SSDs, but TLC with the same correction algorithms will have much higher effective endurance. Performance of QLC drives is currently an unknown quantity. Though many believe they can't possibly be slower than spinning disks (and they are likely correct), the first jmicron based SSDs were significantly slower than spinning disks, so it is possible that performance won't meet all expectations either. There is also a question of unpowered data retention and the effects of storing infrequently accessed data (periodic rewrites?, data retention issues?).
MrSpadge - Tuesday, February 13, 2018 - link
More important than the flash type is a useful endurance rating, which also includes the effect of the controller. A drive with twice as many P/E cycles doesn't help you if the controller causes 3 times the write amplification. Or: QLC with 1000 P/E cycles is more durable than TLC on an older controller with e.g. 800 P/E cycles.BurntMyBacon - Thursday, February 15, 2018 - link
True, but enterprise might prefer the TLC with the newer(better) controller and 1600 to 2000 P/E cycles.DanNeely - Monday, February 12, 2018 - link
Seeing that 1000 P/E cycle number I'm torn between hoping it's only marginally slower than TLC to minimize the performance hit when it inevitably takes over value SSDs, or hoping that its performance is so much worse than TLC that it's not viable for a general purpose SSD.Flunk - Monday, February 12, 2018 - link
Can we start calling multi-level cell NAND dual-level cell NAND yet? Calling it multi-level doesn't make any sense with triple and quad-level NAND around.Hul8 - Monday, February 12, 2018 - link
Then people will get confused why you're bringing up game downloadable content (DLC) in a text about non-volatile memory. :-)takeshi7 - Monday, February 12, 2018 - link
Why not call it BLC for Bi-level cell. Then it won't get confused with DLC.close - Tuesday, February 13, 2018 - link
Then they'll assume MLC is mono-level cell.When it comes to marketing people any good argument will fly over their heads.
saratoga4 - Monday, February 12, 2018 - link
>Calling it multi-level doesn't make any sense with triple and quad-level NAND around.None of the name make any sense, since "triple level" actually has 8 levels, and "quad level" actually has 16 levels. They should have just gone with "2/3/4 bit per cell".
shabby - Monday, February 12, 2018 - link
Everyone knows mlc is dual level nand, i mean no one in their right mind would try to pass off qlc as mlc... right?Holliday75 - Monday, February 12, 2018 - link
I am getting a headache from reading these posts.Kristian Vättö - Tuesday, February 13, 2018 - link
Some vendors are using 3bit and 4bit MLC names.danielfranklin - Monday, February 12, 2018 - link
Does the thought of these QLC chips bring on a shudder of fear in anyone else?Yes they are starting at enterprise and thats fine, but you know these are going to end up in consumer SSDs, the Micron chips always do.
This is not going to be understood by 98% of the market.
OEMs will fill their laptops with them, retailers will be selling them at competitive prices.
We saw so much bad TLC flood the market early on and while i give it to them that its got a lot better, anyone who knows quite how bad QLC will be should be quite worried.
I guess in a flash filled future there are simply going to be many grades of it in existence, unfortunately we just too often see a race to the bottom....
dromoxen - Tuesday, February 13, 2018 - link
These do not fill me with dread , as long as they can keep the write amplification down, a feat I think they are perfecting from knowledge of 3bit cell. . Chips are getting smaller, and cheaper so yes all flash desktop PC, for everyone, should be here in 2020FunBunny2 - Tuesday, February 13, 2018 - link
"Chips are getting smaller, and cheaper "recall that old dodge: "if you step halfway to the wall, how many steps does it take to get to the wall?" we're within cite of Heisenberg NAND. not something I'm thrilled about.
PeachNCream - Tuesday, February 13, 2018 - link
Like Heisenberg blimp NAND? Oh the huge manatee!BurntMyBacon - Thursday, February 15, 2018 - link
@dromoxen: "Chips are getting smaller, and cheaper so yes all flash desktop PC, for everyone, should be here in 2020"Do you really think a ~20% price reduction at the same capacity vs a TLC drive is going to be the tipping point here? I think better stacking, which helps TLC and even MLC as well, is playing a bigger role here.
CheapSushi - Tuesday, February 13, 2018 - link
Heck no, I'm excited and looking forward to QLC SATA/AHCI for bulk storage. I'm waiting on drives to come out to start my homelab builds. Hell, I want them to be DRAMless also, to make BoM cheaper. We've got plenty of NVMe/PCIe drives now for cache & performance including improved V-NAND TLC (which lasts MUCH longer than you'd probably give credit to, upcoming Z-NAND (which acts like SLC) and also Optane (which I think is GREAT; looking forward to x4 laned M.2s). I see zero issue with QLC in laptops and other products. I'm guessing you're the type of person to only use or want a SINGLE drive for an entire system? Ehhh. I guess I see issue there then.CheapSushi - Tuesday, February 13, 2018 - link
Forgot to add that a QLC NAND drive will still be better at random read & write, and bursts, than any HDD, 3.5" or 2.5" that budget products would have. And that's a performance region where the vast majority of said systems are working in. High queue depth, sequential, sustained operations are rarer.BurntMyBacon - Thursday, February 15, 2018 - link
@CheapSushi: "Forgot to add that a QLC NAND drive will still be better at random read & write, and bursts, than any HDD, 3.5" or 2.5" that budget products would have. And that's a performance region where the vast majority of said systems are working in."I believe these will be targeted at what HDDs are currently targeted at. What you state is true if these are used as a primary storage device on a system that would otherwise have a HDD filling that role. However, in a secondary bulk storage role where there is a much lower frequency of access and accesses are much more sequential in nature, the superior random access characteristics may go unnoticed.
danielfranklin - Tuesday, February 13, 2018 - link
My concern is certainly not for me or you, but for the millions of consumers out there who do only run a single disk system.I fully agree in a a perfect world more levels of nand for people in the know is better, ive just seen how the market deals with this before and thats where my concern is.
theuglyman0war - Tuesday, February 13, 2018 - link
Not read anything on what impact the new pci-e 4.0 and upcoming 2020 pci-e 5.0 versions will have on storage performance? New sata specs to leverage as much? As well as Product lines in the works that take advantage of the new versions?( been stuck at pci-e 2.0 buying into x58...
Incremental performance iterations since then has soured my enthusiasm for upgrading past my i7-980x for which remains quite adequate for my workstation needs .vrt pulling in zbrush...
But since Coffee-lake introduced 6 cores providing an obviously sexy hedt alternative that matches my current core count... The upgrade enthusiasm is back )
However...
I am in no hurry to buy into a defunkt PCI-E version again and suffer storage compromises ( like I have under marvell's oppressive sata 3 controller on the x58 for the past 8 years! On the same dam year PCI-E 3.0 released! Talk about bad timing )
Personally hoping the rumored 8-core Cannon-lake releases in April ( further blurring hedt/xeon skus ) with PCI-E 4.0 implemented in motherboard lines hopefully in addition to a slew of announcements from all the storage players touting new tech/product lines leveraging the bandwidth/thruput?
Where PCI-E 5.0 scheduled in a year and a half
theuglyman0war - Tuesday, February 13, 2018 - link
Sure seems making a major investment risky? After 8 years of 3.0...I would hate to be stuck waiting out another 8 years stuck at 4.0 because I jumped the gun in this clausterphobic release season right before 5.0 which again supposedly doubles bandwidth?
( would be like being stuck at 3.0 while tje rest of the world got to realize a comparably NVLINK advantage )
Ghetto resentment!
Trying to not get to excited for anything and just make tmp builds till 5.0...
Just seems to important of a timing concern to make major investments without consideration of 5.0 128GBps lanes compared to today's 32GBps or even stuck at 4.0's 64GBps bandwidth waxing for that 2x leap.
frenchy_2001 - Wednesday, February 14, 2018 - link
PCIe 4.0 is not even implemented yet and you're already talking about 5.0? 0_oI think you are putting too much faith in the industry if you think it will move that fast.
I understand you do not want to "miss out" on the latest tech, but I feel you are misguided.
If you look back at PCIe 2.0 to 3.0, very few applications actually use it.
Go check benchmarks of recent graphic cards tested on Sandy Bridge and latest processors to see that your loss is more theoretical than real.
Storage has moved from SATA3 to PCIe, but even then, the real usage difference between the best SATA3 SSD (which you can still use on your platform) and a more recent PCIe 3.0 SSD is negligible (unless you're running a storage heavy server load).
Conclusion: Upgrade when you need to, to an equipment you'll be happy with, not when the technology changes (unless you know a new product is literally right around the corner in a matter of weeks). The march of technology never stops and there is *always* something better coming (intel and nvidia have new products every year after all).
FunBunny2 - Tuesday, February 13, 2018 - link
whether the P/E value is credible will be determined by the node used. current TLC on 3D is said to be on 40-50 nm, thus gaining charge count/cell. one might hope that QLC is at least on these nodes.frenchy_2001 - Wednesday, February 14, 2018 - link
There was no plan to shrink the node for 3D NAND last I heard.See 3D NAND as a tower. They need a "base" of a certain size, otherwise, they risk it toppling over.
So far, they just increase the number of floors to increase density (from 32 layers to 64 then 96).