SanDisk announces 4TB SSD, hopes for 8TB next year.

http://www.computerworld.com/s/article/print/9248070/SanDisk_announces_4TB_SSD_hopes_for_8TB_next_year

by sc90 • 12 years ago

215 148 12 years ago

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)

Nux• 12 years ago

You'd think; but with all this "centralisation" trends (fb/g+/cloud/etc) there might be no need for them.

phaemon2• 12 years ago

Everyone will need one to hold the blockchain ;)

rodgerd• 12 years ago

It's not the first wave of centralisation, and it won't be the last.

phaemon2• 12 years ago

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).

[Edit: Yes, thanks, I meant TB, not GB. Fixed.]

gberger• 12 years ago

You mean 1 TB, right?

mantrax5• 12 years ago

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.

listic2• 12 years ago

I think we can't count on Moore's law anymore http://en.wikipedia.org/wiki/5_nanometer

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.

amund• 12 years ago

Perhaps one will get a combination of ssd and DNA storage eventually? SSD for low latency storage and DNA for petabytes. Ref 700 TB per gram DNA http://www.extremetech.com/extreme/134672-harvard-cracks-dna...

lazyjones• 12 years ago

15 years? I'd say no longer than 25 years

Sounds reasonable, considering that we were at 1GB (HDD in laptops) 20 years ago.

Expez• 12 years ago

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?

31reasons• 12 years ago

Then why the entry level Mac Book Air still has paltry 250 GB! I hope they catchup with the rest of the world soon.

nixgeek• 12 years ago

These new drives from SanDisk are still over $1/GB in cost.

kylec2• 12 years ago

The entry level MacBook Airs (and retina MacBook Pros) only have 128GB SSDs. I get that Apple wants to keep the price down so they can keep their margins high, but it's a bit small for a primary computer in 2014.

Tloewald• 12 years ago

These hard drives are enterprise targeted and, I assume, rather expensive. Microsoft was selling Surface Pros with 32 GB not long ago, with over half occupied by Windows.

jmpe• 12 years ago

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.

ChuckMcM2• 12 years ago

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.

jacquesm5• 12 years ago

This is a precursor to the next revolution in chips, 3D. It's funny how chips have of course always been three dimensional but the heat removal issues have so far stopped true 3D from happening (as far as I know). But SSDs have a completely different use pattern and likely don't run nearly as many Watts per volume of silicon so you could probably stack them way higher even if you didn't reduce the thickness of the layers all that much. Then it becomes a matter of routing and placement rather than heat dissipation.

Xilinx is already doing something like this: https://www.semiwiki.com/forum/showwiki.php?title=Semi+Wiki:...

Eventually this might even converge on something like a spherical shape, that's when the sum length of the connections is minimal (the surface area of the sphere is too, and that again is a heat removal problem).

sitkack6• 12 years ago

I'd use RF from the controller to the different layers then you would only have to run power. Could even have 2 or 3 transmitters in phased array with little islands of storage on layers of silicon. This would also enable redundant storage (multiple recv) and simultaneous read and write (mimo). The distances in are cm and the signals wouldn't leave the drive package. With each layer exposing power and ground to the layer above the architecture would be scalable in Z just by adding more independent wafers.

https://eeweb.ee.ucla.edu/publications/journalVwaniRoychowdh...

http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.159...

samstave• 12 years ago

Intel has been working on voxel, light-based vertical interconnects for nearly a decade.

I recall when I was there in '99 talking with people about stacking cores. Then a few years later some friends there were doing 64 core research in ~2003-4-ish...

I'm sure they have evolved the voxels since then.

velodrome2• 12 years ago

Here is a paper from two years ago talking about the the theoretical limits of SSD technology.

The Bleak Future of NAND Flash Memory: http://cseweb.ucsd.edu/users/swanson/papers/FAST2012BleakFla...

Leynos• 12 years ago

The next doubling in capacity will likely come about through 3D NAND which allows for much larger feature sizes than current planar NAND. Samsung already have 3D NAND on the market (in 128Gb / die capacity), and other manufacturers intend to go down this route too.

jhallenworld4• 12 years ago

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."

lomnakkus• 12 years ago

"I'm sure they wear out quick with heavy writes."

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.

jhallenworld4• 12 years ago

Well here is one, but this is really for the device level: http://nepp.nasa.gov/workshops/etw2012/talks/Tuesday/T04_Hei...

If I find a good system-level paper, I'll post it.

Anyway, SSD drive wear tends to be rated like this: 5 years for x number of full drive writes per day. For good drives, x is 1. For really good SLC NAND based drives x could be as much as 10, but they are very expensive. For cheap consumer drives x could be 1/10. Based on vendor sales meetings, the aggressively large drives also have low write endurance (but then they say they are not optimized for heavy writes).

Better drives should have a built-in write counter so you can track the remaining life.

Another interesting parameter is their power-off "data retention" time. At end of life, this tends to be specified to be just 90 days.

tim3332• 12 years ago

There seems a lot of contradictory data out there but Techreport did an experiment with 6 consumer grade 256GB drives endlessly reading and writing:

http://techreport.com/review/24841/introducing-the-ssd-endur...

They were still going after 600TB of read writes though with some degradation:

http://techreport.com/review/26058/the-ssd-endurance-experim...

colechristensen• 12 years ago

How many more movies can Netflix support with these new drives?

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.

kordless• 12 years ago

AWS' monthly costs approach cost of ownership. Netflix's costs in providing service are a substantial amount of their spend. AWS works well for Netflix because they have a workload that varies on a daily and weekly basis - by using AWS' excess capacity during peak times, Netflix is able to save money over owning their infrastructure outright.

Given AWS is one big Tetris game, jamming instances in where they fit, this has worked out well for both parties. Until now, with 'now' being defined as the dawn of distributed commodity compute.

latch• 12 years ago

The AWS marketing folks don't get paid enough.

https://www.netflix.com/openconnect/hardware

insaneirish• 12 years ago

This is untrue. AWS is used for some things, but not all.

I've heard certain details I'm unwilling to share, but some of what they do is public (https://www.netflix.com/openconnect/hardware).

Also, they pay developers to work on optimizing network performance of FreeBSD for this custom hardware.

jedberg• 12 years ago

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.

jhallenworld4• 12 years ago

Hmm.. I read that a 2 hour Netflix movie is about 2 GB. So you need ~2.2 Mbps per user or 183 iops / client in terms of 1.5K packets.

One drive can fit 2000 movies. 75K iops/drive / 183 iops/client = 400 clients/drive (this is 900 Mbps, but the drive says up to 400 MB/s, so not bandwidth limited).

Put 16 in a 2U server for 6400 clients. This is 14 Gbps, so you need two 10 G NICs to complete our Netflix appliance.

But you are right, if caching works we don't need all these drives. With 128 GB of RAM we can fit 64 movies in RAM. I could easily believe that most people only watch the popular movies.

reitzensteinm2• 12 years ago

Some quick figures to put it in perspective:

According to instantwatcher.com, there are 6843 movies on Netflix streaming currently. If we conservatively say that the 1080p stream is 2gb, and the other formats add up to another 2gb, that's a total of ~27 TB of space required.

There are 3657 seasons of TV, let's say an average of 2gb per episode, and an average of 14 episodes a season. That's ~102 TB.

So napkin math suggests ~129TB of storage to hold all of Netflix, or 33 of these drives.

Since they need redundancy and to have servers distributed around the country for peering, and these drives will probably cost ~$10k each (5x premium on consumer), I doubt they'll be putting their long tail content on them.

How many people in Boston are currently watching Mission Impossible, the TV series from 1966? Or content like it? A standard hard drive can serve 500 people HD streams so long as you cache ahead into memory intelligently in big enough chunks.

Within a few generations, SSDs will probably be cheap enough to not even bother with mechanical, for this and many other use cases.

[1]http://instantwatcher.com/titles/all?tv_filter=movies+only

zatkin• 12 years ago

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.

virtuallynathan• 12 years ago

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.

bananas2• 12 years ago

It'll be 2022 before I can afford one though... (I still have a 128Gb SSD).

zanny2• 12 years ago

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.

sitkack6• 12 years ago

3.5 years for commodity 4TB SSD or less. I have a 1TB SSD, would gladly switch to a 2TB model for twice the price. As the Sandisk announcement shows, it is market not physical forces that will get us to 4TB.

mikhailt5• 12 years ago

I foresee two types for popular SSDs.

1. MLC at 512GB-1TB is all that needed for the boot/app drives for the next 5 years. They're already at decent prices, 3D stacked chips and another smaller process should be possible within 5 years to bring 1TB to about 200$. 2. TLC at 4TB+ for media/archival storage. I'd foresee 2TB TLC drive at 100$ within 5 years.

neom• 12 years ago

I wish intel was innovating a little more.

listic2• 12 years ago

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.

agumonkey• 12 years ago

And they already have a money bleeding segment : smartphone class SoC.

wmf• 12 years ago

Intel makes the best enterprise SSDs, but they abandoned the consumer SSD market.

rincebrain2• 12 years ago

By what standard did they abandon the consumer market? Their 530 drives come in all the standard consumer form factors and their 330/335 drives still work wonderfully.

coreymgilmore• 12 years ago

Is it just me, or are the performance numbers a bit disappointing given the size and price tag? I guess its a balance between size and speed.

Igglyboo2• 12 years ago

The drives are for enterprise datacenters, not consumers.

rasz_pl• 12 years ago

Thats a cop out. Its synonymous with 'yes they suck, but we will charge arm and a leg for them so some enterprise sucker is bound to bite'.

They also have VERY limited writes/sector count (brobably in the hundreds) before they will start to fail, this has been handled by another PR statement "The drive is aimed at read-intensive applications"

dredmorbius• 12 years ago

Solid throughput or random IOPS?

Latter looks pretty good to me, particularly if you're used to spinning rust.

mkempe2• 12 years ago

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.

adventured2• 12 years ago

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.

nardi4• 12 years ago

That already happened. The biggest computer manufacturer (Apple) doesn't sell a machine without an SSD.

redial2• 12 years ago

http://store.apple.com/us/buy-mac/imac

I think you are being a little premature. All iMacs have hard drives by default.

konstruktor• 12 years ago

Interesting and eye-opening statement, but not entirely correct: The entry level iMac and Mac Mini SKUs have regular hard drives.

mcintyre19942• 12 years ago

Are Apple only the biggest if we include iPhones and/or iPads? I don't think anybody puts an HDD into a smartphone/tablet (it's surely all flash?) so it seems we'd have to limit the definition of computer to laptop/desktop to get anything interesting. Otherwise machines with an HDD are a minority already like you said.

mikhailt5• 12 years ago

Apple is not the biggest manufacturer, it's the most profitable but it is not the biggest one by far.

http://appleinsider.com/articles/14/01/09/apples-domestic-ma...

edward3• 12 years ago

See my website that tracks price per TB of HDD and SSD:

http://edwardbetts.com/price_per_tb/

mkempe2• 12 years ago

That's very nice. Are you keeping historical data? These would be interesting as diagrams.

Carioca• 12 years ago

You have laptop batteries in your SSD listings

beachstartup• 12 years ago

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).

userbinator• 12 years ago

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.

d0ugie• 12 years ago

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?

SanDisk announces 4TB SSD, hopes for 8TB next year.

Related Stories