Thinking about 96-Layer 3D NAND just makes my head hurt. I can't believe the level of complexity we have arrived at to make NAND or for that matter the crazy things we have to do to use EUV to make the latest CPU's. The insane level of engineering going into computers today just blows my mind.
What's important now is: where does it end? Are thousands or tens of thousands of layers doable without increasing die/package size too much?
Hopefully, NAND will be replaced by a universal memory technology before we reach that point. Then all the space formerly used by DRAM and NAND can be occupied by universal memory. The universal memory could also be integrated into the CPU (3DSoC).
Forget about Universal Memory, that is a pipe dream, you will always have trade offs, From Bandwidth, Latency, Capacity, Power, and Cost.
NAND could definitely see 512 / 1024 layers, so $50 / 1TB may only be a few years off. After that we don't know. Do we continue to optimise for Capacity ( which will drive down cost, something manufactures don't want ), or what?
Hopefully where this is all going is price parity (or better) with mechanical hard drives, especially for WORM workloads. There seems to be a reasonable chance of this since $/GB has been pretty stagnant for many years in the mechanical hard drive market.
I can't presume to understand all of the factors, but based on what the industry has done it seems to be true that more layers leads to lower cost per bit. Early on we saw the transition between planar NAND, which scaled with traditional lithography nodes (smaller features leading to higher density NAND). The transition to 3D NAND accelerated the reduction trend of cost per bit and improved durability thanks to being able to fall back to larger feature sizes. Further increases in layers have led to further reductions of cost per bit. While adding more layers adds certain costs, like more CVD machines, they are fixed costs, which will be amortized over the life of the factory and over thousands of wafers. So while costs are higher, they are evidently less proportionately than the increase in bits per wafer.
The continued reductions in cost per bit are not being driven entirely or even predominantly by the increasing of layer count. That's already well beyond the point of diminishing returns, and increasing the layer count on its own really does increase costs by about the same amount that it increases density. The NAND manufacturers are working their way through a whole bag of other tricks to increase density and reduce cost, and in a lot of cases having a higher layer count means those other changes have a bigger impact.
When I can buy 8TB SSDs for about $200 we'll be there. Currently, if they were even widely available and using your pricing, an 8TB SSD would cost $680 (I'm assuming USD). That's asking for an almost a 3.5x price drop in 3-4 years, that's nothing to sneeze at.
It'll get there eventually. That's why I question the wisdom of continuing hard drive production at all. I've had tons of problems with hard drives, but no problems with SSDs dating all the way back to the first gen 120 gb drives.
The one thing that is going to hinder lower prices is a lower adoption rate due to decreased failures. For instance, Seagate is widely known as a budget brand for hard drives. Their drives are often among the cheapest around, but they can have high failure rates depending on the drive/capacity. At one company, we added a 48 bay NAS with 120gb SSDs in RAID, and there hasn't been a single drive failure yet. Compare that to the previous NAS that had a bunch of spinning drives and you had a drive fail every 6-18 months. The NAS in question has been in use for MANY years now and it wasn't cheap to put together, but the performance impact was worth it. The 10 Gbit link to the graphics editors from the rack means that they are able to pull down graphics and video quite fast.
With that kind of reliability, the floor for SSDs will always be higher than HDDs. There isn't any getting around it. Even if you warrantied your defective hard drives, you still had to buy a new one or keep a few spare on hand just so you could get that RAID rebuilt ASAP. With hard drives, most companies wouldn't bother with the RMA due to the time it takes to RMA the drive, send it in, track the progress, and get it back because it often costed more than simply purchasing a new drive.
Come to think of it, I don't recall a single company I ever worked at that RMAed a hard drive. I have personally, but never at a company. We would just pop in a new one and send the defective one to the shredder.
well problem grows faster than linear so 512 might end up with problems we dont even dream about now. Overheat of "middle layers" or something that is bizarre right now, with 1024 stacks will happen. I think we will stop on ~384'ish per stack. QUOTING: "What is important is that Samsung plans to use its 136-layer architecture with speed-optimized circuit design to build V-NAND devices with over 300 layers by mounting three of the current stacks on top of each other (thus tripling a chip’s capacity)." So it will end up as ~3x384'ish per die ~700 Gb so 3 chips for 256 GB drive... as current best 256 Gb chips cost ~60$ in final product, we can expect 1TB in ~70-80$ brackets in ~1.5 years.
It doesn't 'drive down cost'. If a 256 gb SSD costed $50 to make in say, 2016 and now the 512 gb costs $50 to make, that isn't 'lost revenue'. Margins aren't driven down by capacity increases, but instead they are driven down by competitive pressure.
I wish Anandtech comment has a tag system so I could tag those who call people idiot, naive, stupid ( to quote their words) for saying NAND and DRAM are price fixing Cartel and market manipulation.
The amount of engineering required in modern day NAND and DRAM, and cost of R&D and Fab Expansion is quite insane.
Well, while not all of this is price fixing, Samsung, for example, has been convicted of NAND price fixing already. Same thing for LCD panels and other things. In fact, several years ago, Apple was going to open a memory plant with Samsung, but after Samsung’s conviction, which resulted in another several hundred million dollar fine, decided not to.
Actually, it was before we had EUV that we had to do really crazy things to make chips that had features smaller than the wavelength of the light used to make them - stuff like quadruple patterning. Making the chips in a vacuum, using shallow reflection off hyperbolic mirrors to focus the EUV light, by comparison, those aren't crazy things, that's just what you have to do to use that kind of light (if you can call it that instead of soft X-rays).
So you don't think using a laser that is the size of 5 refrigerators to generate a high intensity laser beam that is strobed low/high over and over to hit tin droplets being fired at 100mph in a vacuum to just create the light is not complex? The low strobe hits the tin to spread the droplet flat then the high one obliterates it to create the EUV. This sounds a tad complex to me. /smh
Doesn't each new layer require additional vertical vias passing through the layers below? I'm imagining this as requiring a dedicated elevator shaft for each layer. Eventually the horizontal footprints of the elevator shafts passing through the lower layers cancels out the extra space gained by adding the next layer.
Awesome news. I don’t work for Micron any more but I went to Singapore and bootstrapped the two dozen or so racks of infrastructure to run the initial Fab 10 green field build, in 2007 or so. I always enjoy hearing about where they are now.
It should be noted that this type of expansion comes at a great cost, its a risky move simply because the investment is so large that to recoup costs will have to please investors. This upgrade can literally ruin Micron. This is why you see big tech companies sell off portions of "underperforming" areas of the companies..they will sell this off to a competitor just to invest in something more sure.
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
21 Comments
Back to Article
FreckledTrout - Friday, August 16, 2019 - link
Thinking about 96-Layer 3D NAND just makes my head hurt. I can't believe the level of complexity we have arrived at to make NAND or for that matter the crazy things we have to do to use EUV to make the latest CPU's. The insane level of engineering going into computers today just blows my mind.quiksilvr - Friday, August 16, 2019 - link
Wrap your brain around this: Samsung is already on 6th generation with 136-Layer 3D NAND.https://www.anandtech.com/show/14721/samsung-unvei...
nandnandnand - Saturday, August 17, 2019 - link
What's important now is: where does it end? Are thousands or tens of thousands of layers doable without increasing die/package size too much?Hopefully, NAND will be replaced by a universal memory technology before we reach that point. Then all the space formerly used by DRAM and NAND can be occupied by universal memory. The universal memory could also be integrated into the CPU (3DSoC).
ksec - Saturday, August 17, 2019 - link
Forget about Universal Memory, that is a pipe dream, you will always have trade offs, From Bandwidth, Latency, Capacity, Power, and Cost.NAND could definitely see 512 / 1024 layers, so $50 / 1TB may only be a few years off.
After that we don't know. Do we continue to optimise for Capacity ( which will drive down cost, something manufactures don't want ), or what?
But that is at least 3 - 4 years off.
3DoubleD - Saturday, August 17, 2019 - link
Hopefully where this is all going is price parity (or better) with mechanical hard drives, especially for WORM workloads. There seems to be a reasonable chance of this since $/GB has been pretty stagnant for many years in the mechanical hard drive market.ajp_anton - Saturday, August 17, 2019 - link
Does the number of layers really lower the prices that much? Silicon area goes down, but costs per area goes up.3DoubleD - Saturday, August 17, 2019 - link
I can't presume to understand all of the factors, but based on what the industry has done it seems to be true that more layers leads to lower cost per bit. Early on we saw the transition between planar NAND, which scaled with traditional lithography nodes (smaller features leading to higher density NAND). The transition to 3D NAND accelerated the reduction trend of cost per bit and improved durability thanks to being able to fall back to larger feature sizes. Further increases in layers have led to further reductions of cost per bit. While adding more layers adds certain costs, like more CVD machines, they are fixed costs, which will be amortized over the life of the factory and over thousands of wafers. So while costs are higher, they are evidently less proportionately than the increase in bits per wafer.Billy Tallis - Sunday, August 18, 2019 - link
The continued reductions in cost per bit are not being driven entirely or even predominantly by the increasing of layer count. That's already well beyond the point of diminishing returns, and increasing the layer count on its own really does increase costs by about the same amount that it increases density. The NAND manufacturers are working their way through a whole bag of other tricks to increase density and reduce cost, and in a lot of cases having a higher layer count means those other changes have a bigger impact.nandnandnand - Saturday, August 17, 2019 - link
Your idea of what is a pipe dream or 3-4 years off is hopelessly warped. We can already find 1 TB for about $85.3DoubleD - Saturday, August 17, 2019 - link
When I can buy 8TB SSDs for about $200 we'll be there. Currently, if they were even widely available and using your pricing, an 8TB SSD would cost $680 (I'm assuming USD). That's asking for an almost a 3.5x price drop in 3-4 years, that's nothing to sneeze at.eek2121 - Monday, August 19, 2019 - link
It'll get there eventually. That's why I question the wisdom of continuing hard drive production at all. I've had tons of problems with hard drives, but no problems with SSDs dating all the way back to the first gen 120 gb drives.The one thing that is going to hinder lower prices is a lower adoption rate due to decreased failures. For instance, Seagate is widely known as a budget brand for hard drives. Their drives are often among the cheapest around, but they can have high failure rates depending on the drive/capacity. At one company, we added a 48 bay NAS with 120gb SSDs in RAID, and there hasn't been a single drive failure yet. Compare that to the previous NAS that had a bunch of spinning drives and you had a drive fail every 6-18 months. The NAS in question has been in use for MANY years now and it wasn't cheap to put together, but the performance impact was worth it. The 10 Gbit link to the graphics editors from the rack means that they are able to pull down graphics and video quite fast.
With that kind of reliability, the floor for SSDs will always be higher than HDDs. There isn't any getting around it. Even if you warrantied your defective hard drives, you still had to buy a new one or keep a few spare on hand just so you could get that RAID rebuilt ASAP. With hard drives, most companies wouldn't bother with the RMA due to the time it takes to RMA the drive, send it in, track the progress, and get it back because it often costed more than simply purchasing a new drive.
Come to think of it, I don't recall a single company I ever worked at that RMAed a hard drive. I have personally, but never at a company. We would just pop in a new one and send the defective one to the shredder.
deil - Monday, August 19, 2019 - link
Well 1TB of ssd != 1TB of FAST ssd. I expect 5GB/s in that drive for 80$ and current one will not reach full sata speeds...deil - Monday, August 19, 2019 - link
well problem grows faster than linear so 512 might end up with problems we dont even dream about now. Overheat of "middle layers" or something that is bizarre right now, with 1024 stacks will happen. I think we will stop on ~384'ish per stack.QUOTING:
"What is important is that Samsung plans to use its 136-layer architecture with speed-optimized circuit design to build V-NAND devices with over 300 layers by mounting three of the current stacks on top of each other (thus tripling a chip’s capacity)."
So it will end up as ~3x384'ish per die ~700 Gb so 3 chips for 256 GB drive...
as current best 256 Gb chips cost ~60$ in final product, we can expect 1TB in ~70-80$ brackets in ~1.5 years.
eek2121 - Monday, August 19, 2019 - link
It doesn't 'drive down cost'. If a 256 gb SSD costed $50 to make in say, 2016 and now the 512 gb costs $50 to make, that isn't 'lost revenue'. Margins aren't driven down by capacity increases, but instead they are driven down by competitive pressure.ksec - Saturday, August 17, 2019 - link
I wish Anandtech comment has a tag system so I could tag those who call people idiot, naive, stupid ( to quote their words) for saying NAND and DRAM are price fixing Cartel and market manipulation.The amount of engineering required in modern day NAND and DRAM, and cost of R&D and Fab Expansion is quite insane.
melgross - Monday, August 19, 2019 - link
Well, while not all of this is price fixing, Samsung, for example, has been convicted of NAND price fixing already. Same thing for LCD panels and other things. In fact, several years ago, Apple was going to open a memory plant with Samsung, but after Samsung’s conviction, which resulted in another several hundred million dollar fine, decided not to.quadibloc - Saturday, August 17, 2019 - link
Actually, it was before we had EUV that we had to do really crazy things to make chips that had features smaller than the wavelength of the light used to make them - stuff like quadruple patterning. Making the chips in a vacuum, using shallow reflection off hyperbolic mirrors to focus the EUV light, by comparison, those aren't crazy things, that's just what you have to do to use that kind of light (if you can call it that instead of soft X-rays).FreckledTrout - Sunday, August 18, 2019 - link
So you don't think using a laser that is the size of 5 refrigerators to generate a high intensity laser beam that is strobed low/high over and over to hit tin droplets being fired at 100mph in a vacuum to just create the light is not complex? The low strobe hits the tin to spread the droplet flat then the high one obliterates it to create the EUV. This sounds a tad complex to me. /smhhttps://www.youtube.com/watch?v=5yTARacBxHI
ironwing - Sunday, August 18, 2019 - link
Doesn't each new layer require additional vertical vias passing through the layers below? I'm imagining this as requiring a dedicated elevator shaft for each layer. Eventually the horizontal footprints of the elevator shafts passing through the lower layers cancels out the extra space gained by adding the next layer.sor - Friday, August 16, 2019 - link
Awesome news. I don’t work for Micron any more but I went to Singapore and bootstrapped the two dozen or so racks of infrastructure to run the initial Fab 10 green field build, in 2007 or so. I always enjoy hearing about where they are now.imaheadcase - Sunday, August 18, 2019 - link
It should be noted that this type of expansion comes at a great cost, its a risky move simply because the investment is so large that to recoup costs will have to please investors. This upgrade can literally ruin Micron. This is why you see big tech companies sell off portions of "underperforming" areas of the companies..they will sell this off to a competitor just to invest in something more sure.