Over the last decade SSDs have increased their storage capacity at an incredible rate. This growth has mostly been fueled by adapting better process technology. I read a while back that the move to ever smaller scales (now 19nm) came at the costs of increased errors rates and that the limiting factor going forward would be error correction. Has this situation changed, or is SanDisk just throwing chips at the capacity problem, along with better controllers, and targeting an enterprise segment willing to pay the price?
One of the interesting changes has been folks have been thinning out silicon wafers. Typically 700 microns there have been examples of working wafers as thin as 5 microns. If you stack 5 micron silicon you can put 100 chips in the same place you put 1 before, assuming you can figure out how to talk between them. Thru-silicon-vias (TSV) is the current poster child but there are better alternatives coming. Should be interesting when that comes to market.
Yes. When scaling down semiconductors your oxide shows weird behavior. In transistors this was one of the crisis that hindered miniaturisation, there were alternatives (like exotic high-k dielectrics) but semi corps tend to steer away from such things.
Miniaturization happens not only on the surface (length and width) but also depth. This is where silicon oxides come into play. Flash tech relies either on charge trapping in the oxide itself or in a floating gate. Both become sensitive to effects like tunneling when you scale them down in the 10nm magnitude.
With all the process shrinks, and packing more bits per cell, NAND flash is getting to the point where its characteristics are closer to DRAM than traditional magnetic storage; like hard drives the data will stay there for a while when they're powered off, but like DRAM, it's not going to be there forever. However, unlike DRAM and more like hard drives, they wear out.
For enterprise cache-like applications this makes sense, but with DRAM prices not that far off (only a few times), I wonder if battery-backed DRAM might actually offer better value (and theoretically could be far higher performing) than having to replace worn-out SSDs periodically.
Fun Saturday question. It's 2014. How long until we see notebooks (or tablets or any other portable computing device) with 1PB of storage?
15 years? I'd say no longer than 25 years -- if we're still carrying around computing devices by then (and it hasn't all just been subsumed into the cloud)
van der Waals' radius of silicon is 0.21 nm http://en.wikipedia.org/wiki/Silicon so current 19nm process nodes are already 45 atoms wide, there simply isn't much room for improvement left.
and it looks even deader for Flash than it does for CPU's, as ever finer processes are already producing ever slower and less reliable Flash cells. Those can be, and certainly are, mitigated by ever more extensive read/write parallelization and error correction in the controller, but only so much. "The drive is aimed at read-intensive applications" is a nice way to say that write performance sucks, relatively speaking.
So, if the semiconductor industry won't come up with something unheard of (which it just might, given the scale of the stakes involved) this might be one of the last radical upgrades available.
If you assume we're on 1 TB at the moment, and capacity doubles every 18 months (SSDs are a bit ahead of this at the moment), then 15 years is about right.
By 2030 seems like a fairly safe bet (barring major disaster before then).
Linear projections don't work. This is why by linear projection we should have 20GHz 128 core CPUs but things have progressed differently in history.
Capacity will grow until it hits one of two types of limits: physics or the market. If physics doesn't allow for 1PB SSD in a notebook, you won't get it. If the market doesn't see the value in 1PB notebook disks (and I'd claim they won't see it, even after 15 years) you won't see notebooks with 1PB SSD.
I'm sure they wear out quick with heavy writes. On the other hand, the killer app for these has to be media streaming. How many more movies can Netflix support with these new drives? I wonder if they are space limited or iops limited.
"The drive is aimed at read-intensive applications, such as data warehousing, media streaming and web servers. The typical workload envisioned for the 4TB drive is 90% read and 10% write, SanDisk stated."
Netflix has 28TB of Flash storage in a 1U box (e3-12xxv2, 125w system power usage) with 40Gbit/s of Network capacity, which gets fully saturated. They run BSD, and Nginx, I believe. This would double their density.
> I wonder if [Netflix is] space limited or iops limited.
I don't think you're asking the right question. The more iops, the less space and the less cost. I'd say Netflix is always optimizing for both.
The right question to ask is what is the break-even point between space and cost where these drives make sense, and that would depend mostly on how popular the most popular content is. If everyone were watching the same few things, but enough that it can't just all fit in RAM, it would make a lot of sense to have a lot of these drives. But if the watching is spread out across a lot of content, then not so much.
Do you have something to back that assertion up? Seriously, I'm interested, so if you have links to further information, I'm sure everyone would appreciate it.
It's important to note that most companies these days that have 'big data' typically follow the WORM (write once, read many) pattern. They utilize technologies that write their data to the disk once, but never delete.
What is the anticipated price/TB curve of SSD over the next five years? at what point would most consumer storage switch from HDD to SSD?
If we start with a HDD/SSD price ratio of 1:7 (based on a quick check with Amazon) and hope for SSDs to get cheaper/TB by 50% each coming 18 months (?) we'd have a strong incentive to switch by year 2019.
I think we're pretty close to the mainstream switching now.
Inexpensive, large external HDDs will fill the gap on media storage for consumers long after the main drive has been switched to SSD for speed.
Consumers aren't going to wait for SSD to catch up to 5tb HDD externals that you'll be able to get for $129. They'll simply buy a system with 512gb to 1tb of SSD storage, and pick up the external if they need it.
just a data point: for our enterprise customers, 800GB intel datacenter 3500 series flash drives (the penultimate model good for most write-heavy applications) are already basically the same price as 600GB 15k SAS drives and actually cost less if you drop the raid controller which isn't that great with ssd (in our experience).
Me too. But I guess Intel is not that much strategically interested in competing in ever more competitive SSD space, with decreasing profit margins and increasingly larger R&D budgets needed. Not that Intel couldn't afford them, but I think it's not as attractive for them as the CPU market right now.
Wow, 8TB, that's a lot of girly pictures! For the same general use purposes, do SSDs put out significantly less heat versus HDDs and also burn less electricity, and enough so to, for example, help eliminate the need for fans?
For the typical consumer, not necessarily what to decide what to stuff into an HP 980, is it now clear that cost per byte is evening out and that soon the typical array of laptops in your local Best Buy including the cheaper of the lot will have SSDs?
Tl;dr, are hard drives being completely phased out faster and faster?
Well, in 2010 a 256GB ssd was upwards of $500+. Today, a TB SSD costs that. So yeah, capacity for price increases 4x over in 4 years. So 2022 is a good estimate for commodity 4TB SSDs.
I wonder if we will ever get nand storage below mechanical in price per GB. Even in 2022 I can image 8 - 10 TB mechanical disks being only $100 still, while the 2TB SSDs cost about that. Because by then we are probably hitting some physical limits.
[+] [-] Expez|12 years ago|reply
[+] [-] ChuckMcM|12 years ago|reply
[+] [-] jmpe|12 years ago|reply
Miniaturization happens not only on the surface (length and width) but also depth. This is where silicon oxides come into play. Flash tech relies either on charge trapping in the oxide itself or in a floating gate. Both become sensitive to effects like tunneling when you scale them down in the 10nm magnitude.
[+] [-] velodrome|12 years ago|reply
The Bleak Future of NAND Flash Memory: http://cseweb.ucsd.edu/users/swanson/papers/FAST2012BleakFla...
[+] [-] 31reasons|12 years ago|reply
[+] [-] userbinator|12 years ago|reply
For enterprise cache-like applications this makes sense, but with DRAM prices not that far off (only a few times), I wonder if battery-backed DRAM might actually offer better value (and theoretically could be far higher performing) than having to replace worn-out SSDs periodically.
[+] [-] DanielBMarkham|12 years ago|reply
15 years? I'd say no longer than 25 years -- if we're still carrying around computing devices by then (and it hasn't all just been subsumed into the cloud)
[+] [-] listic|12 years ago|reply
van der Waals' radius of silicon is 0.21 nm http://en.wikipedia.org/wiki/Silicon so current 19nm process nodes are already 45 atoms wide, there simply isn't much room for improvement left.
and it looks even deader for Flash than it does for CPU's, as ever finer processes are already producing ever slower and less reliable Flash cells. Those can be, and certainly are, mitigated by ever more extensive read/write parallelization and error correction in the controller, but only so much. "The drive is aimed at read-intensive applications" is a nice way to say that write performance sucks, relatively speaking.
So, if the semiconductor industry won't come up with something unheard of (which it just might, given the scale of the stakes involved) this might be one of the last radical upgrades available.
[+] [-] Nux|12 years ago|reply
[+] [-] phaemon|12 years ago|reply
By 2030 seems like a fairly safe bet (barring major disaster before then).
[Edit: Yes, thanks, I meant TB, not GB. Fixed.]
[+] [-] amund|12 years ago|reply
[+] [-] lazyjones|12 years ago|reply
Sounds reasonable, considering that we were at 1GB (HDD in laptops) 20 years ago.
[+] [-] mantrax5|12 years ago|reply
Capacity will grow until it hits one of two types of limits: physics or the market. If physics doesn't allow for 1PB SSD in a notebook, you won't get it. If the market doesn't see the value in 1PB notebook disks (and I'd claim they won't see it, even after 15 years) you won't see notebooks with 1PB SSD.
[+] [-] jhallenworld|12 years ago|reply
"The drive is aimed at read-intensive applications, such as data warehousing, media streaming and web servers. The typical workload envisioned for the 4TB drive is 90% read and 10% write, SanDisk stated."
[+] [-] virtuallynathan|12 years ago|reply
[+] [-] jedberg|12 years ago|reply
I don't think you're asking the right question. The more iops, the less space and the less cost. I'd say Netflix is always optimizing for both.
The right question to ask is what is the break-even point between space and cost where these drives make sense, and that would depend mostly on how popular the most popular content is. If everyone were watching the same few things, but enough that it can't just all fit in RAM, it would make a lot of sense to have a lot of these drives. But if the watching is spread out across a lot of content, then not so much.
[+] [-] lomnakkus|12 years ago|reply
Do you have something to back that assertion up? Seriously, I'm interested, so if you have links to further information, I'm sure everyone would appreciate it.
[+] [-] zatkin|12 years ago|reply
[+] [-] colechristensen|12 years ago|reply
Netflix doesn't do their own hardware, they use AWS. One can also assume that the cost of the content far outweighs the cost of delivery.
[+] [-] mkempe|12 years ago|reply
If we start with a HDD/SSD price ratio of 1:7 (based on a quick check with Amazon) and hope for SSDs to get cheaper/TB by 50% each coming 18 months (?) we'd have a strong incentive to switch by year 2019.
[+] [-] adventured|12 years ago|reply
Inexpensive, large external HDDs will fill the gap on media storage for consumers long after the main drive has been switched to SSD for speed.
Consumers aren't going to wait for SSD to catch up to 5tb HDD externals that you'll be able to get for $129. They'll simply buy a system with 512gb to 1tb of SSD storage, and pick up the external if they need it.
[+] [-] nardi|12 years ago|reply
[+] [-] edward|12 years ago|reply
http://edwardbetts.com/price_per_tb/
[+] [-] beachstartup|12 years ago|reply
[+] [-] neom|12 years ago|reply
[+] [-] listic|12 years ago|reply
[+] [-] wmf|12 years ago|reply
[+] [-] coreymgilmore|12 years ago|reply
[+] [-] Igglyboo|12 years ago|reply
[+] [-] dredmorbius|12 years ago|reply
Latter looks pretty good to me, particularly if you're used to spinning rust.
[+] [-] d0ugie|12 years ago|reply
For the typical consumer, not necessarily what to decide what to stuff into an HP 980, is it now clear that cost per byte is evening out and that soon the typical array of laptops in your local Best Buy including the cheaper of the lot will have SSDs?
Tl;dr, are hard drives being completely phased out faster and faster?
[+] [-] bananas|12 years ago|reply
[+] [-] zanny|12 years ago|reply
I wonder if we will ever get nand storage below mechanical in price per GB. Even in 2022 I can image 8 - 10 TB mechanical disks being only $100 still, while the 2TB SSDs cost about that. Because by then we are probably hitting some physical limits.