Hynix 48-GB Flash MCP

Hal_Porter writes to let us know that the third-largest NAND chip maker, Hynix, has announced they have stacked 24 flash chips in a 1.4mm thick multi-chip package. It's not entirely clear from the article whether the resulting 48-GB device is a proof of concept or a product. The article extrapolates to 384 GB of storage in a single package, sometime. Hal_Porter adds: "It's not clear if it's possible to write to them in parallel — if so the device should be pretty damn fast. The usual objection to NAND flash as a hard drive replacement is lifetime. NAND sectors can only be written 100,000 times or so before they wear out, but wear leveling can be done to spread writes evenly over at least each chip. I worked out that the lifetime should be much longer than a typical magnetic hard disk. There's no information on costs yet frankly and it sounds like an expensive proof of concept, but it shows you the sort of device that will take over from small hard disks in the next few years."

Database servers

[gnuman99 (746007)]

Random seek is probably one of the biggest bottlenecks in large databases. There are even databases that optimize reads/writes to be more consecutive on the disk. A drive like that would throw that problem out of the window.

Re:

[Jeppe Salvesen (101622)]

Well, I'm rooting for this solution for an index tablespace (or .MYI on a MyISAM database), and a regular raid for the actual data. I think that combo would fly, as the tree operations on the index thrives on low latency - and the consecutive reads/writes of table data thrives on regular disk devices as they tend to have a good throughput.

Re:Database servers

[CastrTroy (595695)]

You'd have to look at how much actual reading or writing to the drive is done by a computer from that era. Currently, hard drive space is really cheap, so we write lots of stuff to the disk, like temp files, log files, swap out programs, and even with some filesystems and operating systems write to the drive every time a file is accessed. A computer from that era wouldn't be writing so much stuff too the hard drive, as hard drives were small and expensive. It would likely only write to the drive when you need a program to save actual human created data, or when you install a new program. Reading would only be done when you start up the computer, a new program, or load a file.

Re:

[Surt (22457)]

http://download.micron.com/pdf/datasheets/flash/na nd/4gb_nand_m40a.pdf [micron.com]
promises data retention of 10 years. I would guess that it will function longer than that, but only if you refresh the data.

Re:

[Xichekolas (908635)]

*sigh* [slashdot.org]

Re:

[Atzanteol (99067)]

But you then have the problem with many databases being much larger than the 48GB listed or even the 384GB.

Let me introduce you to our friend RAID [wikipedia.org].

48 GB = 384Gb

[sirket (60694)]

The article does not extrapolate to 384 GB of storage- they extrapolate to 384 Gb of storage which is 48 GB of storage. bits != bytes.

Re:

[sirket (60694)]

JHust to clarify- the company mentions possibly going to 28 stacked chips which would be 448 gigabits (not gigabytes) of storage- or about 56 GB of space. Now as flash chips grow in size- this could double (assuming 32 Gb NAND chips which are becoming available) to 96 or 112 GB of storage or more (assuming larger chips).

Re:48 GB = 384Gb

[timeOday (582209)]

But these appear to be be tiny (for mobile applications).... you could fit an enormous number of them into a 3.5" (or even 2.5") hard drive enclosure, if you can afford it. Put in a controller that can read and write to, say, 16 chips in parallel, and you would have a monster hard drive in every respect.

Re:

[HiThere (15173)]

But what they *ought* to do is figure out how to stack 33 or 35 planes (1 word + parity). I suppose that they could do fancier error correction...but this is a ram chip, not a computer.

Anyway, then if they could read/write all planes in parallel you'd not only have fast access, but also simple addressing. (I.e., you could reasonably do I/O to a single column...admittedly slower than block transfer, but nicer if you only need to change one word.) This would be more important if memory usage cycles were mor

Re:

[steelfood (895457)]

I don't know which K you're using, but in my world, there are 1000 bits per Kb, 8 bits per byte, and 1024 bytes per KB.

Which means 384 billion bits is 48 billion bytes, which is only 44.7GB.

HDD manufacturers want 1000 bytes per KB, but I don't buy that at all. It no different from the ram manufacturers rounding up 536866816 to 512MB when 512MB is actually 536870912.

48GB?

[Tiroth (95112)]

Looks like people are confusing bits and bytes. 48GB does not appear in the article anywhere, so I assume this is obtained by dividing 384Gb by 8.

24 layers x 16Gb package = 384Gb, so the article itself is consistent.

swap space / tmpfs / cacheing

[lobiusmoop (305328)]

Given the low price of RAM these days (1 or 2 gigs being standard) minimising the need for swapping, and availability of tmpfs in the Linux kernel, I'm surprised there are not more flashdrive based linux boxes available these days.

Re:

[Jeff DeMaagd (2015)]

Have you actually bought a sizeable flash drive? 4GB CD cards are starting to be common, I think CF cards are the most affordable flash drive that you can reasonably use as a system drive. But for the same price, you might buy a 300GB hard drive. Not only that, there doesn't seem to be any affordable SATA-based flash drives, which is quickly becoming the only drive connection type found in computers.

So it would work great for a network terminal, there doesn't seem to be enough for most people to use just

Re:

[Com2Kid (142006)]

2GB SD cards are still a better band for your buck, typically. In the very least, compatibility is better. :)

You can get them pretty easily for $20 a pop.

Amazingly enough Amazon has 2GB SD cards cheaper than Newegg. $15 a pop (no free shipping though!)

That is $30 for 4GB, or $60 for 8GB.

Not quite enough to get Vista up and running, but it should do fine for a stand alone Linux box. :-D

I wonder what the throughput would be if a proper hardware controller was put in place and you had 50 of those things in p

Re:

[Spokehedz (599285)]

I can't find it now, but I remember a device that would take a bunch of SD cards (like, 4 slots) and would combine them into a big disk that had (I believe) SATA on it. So, you would take a bunch of these cheap 2GB SD cards, and it would make one big disk out of them all.

http://www.geekstuff4u.com/product_info.php?manufa cturers_id=&products_id=492 [geekstuff4u.com]

Not it, but close. Also way too expensive.

Re:

[by Anonymous Coward]

I'm surprised there are not more flashdrive based linux boxes available these days.

There will be several million shortly...

# Mass storage: 1024 MiB SLC NAND flash, high-speed flash controller;
# Drives: No rotating media.

From the OLPC Spec [laptop.org]

Flash lifespan in persective

[G4from128k (686170)]

Even at only 1,000 writes of reliable lifespan, 48 GB could handle 48 TB of writes or over 4,000 hours of continuous writing of compressed HD video (or about 2 years of 40 hr/week writes of a video stream). Checking my average usage of disk I/O finds that I only average about 2 GB of writes per day which would suggest that this device would last me 24,000 days (or 65 years). And if the life is 10,000 or 100,000, then I'd see 10X or 100X that lifespan.

Your mileage may vary, but I'd bet that 99% of users would never keep their computer (especially a laptop that is the more likely application for flash-based drives) for long enough to see the disk fail from wear.

Re:

[smallfries (601545)]

You're assuming that the 2GB a day could be spread evenly over the disk. This would vary depending on how much free space you have on the device. If your drive is 1% full then you can distribute your writes over the other 99%. But most people don't keep their storage mainly empty. In fact people tend to run just under the limit - hence the saying that crap always expands to fill the available space. If your drive was 99% full then you can't distribute the writes over the parts with data (as it would have to

Are they really run near the limit?

[tepples (727027)]

If your drive is 1% full then you can distribute your writes over the other 99%. But most people don't keep their storage mainly empty. In fact people tend to run just under the limit

Citation needed, at least for common uses of flash memory. One common use case for flash memory is in digital cameras. A photographer shoots a "roll", copies everything from the pictures folder on the flash card to a larger drive, and deletes the "roll" from the flash card. Even for larger drives such as hard disk drives, Windows encourages the user to keep 15 percent of the drive free so that Defrag can work more efficiently.

Re:

[ivan256 (17499)]

You're assuming that the entire capacity of the chip would be exposed to the end user, and none would be reserved for dynamic load leveling.

You're also assuming that unchanged data would never be moved by the load leveling algorithm.

I don't think either are valid assumptions, and you're just plain wrong.

Re:

[msgtomatt (1147195)]

Your calculation of 24,000 days is when the drive reaches total failure. Your logic does apply to camcorder applications in which data is always written sequentially. But, in PC applications you do not write the information as a bit stream, you write things fairly randomly. When you change the contents of a file without changing the file size, you update the same physical memory locations. So after you update your file a 1,000 times, it becomes corrupted and you loose your data. Once a single byte becomes c

Re:

[russotto (537200)]

Unfortunately, wear leveling also means that when one cell fails, many others are likely to go real soon now. So when you start getting bad sectors, time to replace the device.

Re:

[compro01 (777531)]

though keep in mind that the 48GB might really be 49 or 50* to provide spare sectors in the same manner hard drives do.

*numbers not necessarily based on any factual information.

1 MiB - 1 MB = wear leveling

[tepples (727027)]

though keep in mind that the 48GB might really be 49 or 50* to provide spare sectors in the same manner hard drives do.

Based on my experience buying CF and SD cards, this is actually where the 4.8 percent difference between a MB and a MiB goes. When you buy, say, a 512 MB memory card, it is actually a 512 MiB (536 MB) memory card where 4.8 percent of the sectors are spared. I've bought three "1 GB" cards, each of which had 1,024 MB available for files, folders, and allocation data.

implications of flashing

[one_who_uses_unix (68992)]

It used to be that there were serious implications if you engaged in flashing, potentially including jail time!

The world has come a long way when any geek can flash thousands of times and not have problems with his hard disk.

IPod

[dazedNconfuzed (154242)]

iPod Touch, meet Hynix 48-GB Flash MCP!

media storage

[Floritard (1058660)]

It is just writing that is limited right? Myself, I'd love to have the space to host all my media, most of which just sits archived on dvd-r. I'd only need to write to the disk once. Seems most people, aside from those who do video production, really only need large amounts of space to serve/store media. Be cool to just keep a 200 gig SATA for regular use and just keep buying these suckers and fillin' them up for all that media. Later, when they're cheap that is.

HyperDrive4

[myspys (204685)]

Something that should be mentioned when talking about these things is HyperDrive4 [hyperossystems.co.uk]

Re:

[Surt (22457)]

Ouch that's expensive. And big (in physical dimensions) compared to:
http://www.computers4sure.com/product.asp?producti d=5623741&affid=10000483 [computers4sure.com]

I guess it may be somewhat faster, but both are approaching the limits of what you can push through a sata interface.

Nice Butt...

[eno2001 (527078)]

...would you really want to buy something from a company named Hynix? At worst it sounds like a Unix that smells like ass. At best it sounds like a bunch of stoned Unix devels.

What about RAID?

[GreatBunzinni (642500)]

Recently, this whole flash drive business has been popping up in the news, with announcements of a whole gob of commercial solid-state drives based on flash technology and the like. Nonetheless, there is a big void in the flash drive world that, at least at first glance, could be easily filled with trivial technology and off the shelf products but no one seems to be paying any attention.

I'm talking about RAID + flash cards.

Flash cards are everywhere and, although their cost per GB is rather high, a 1GB card is easily affordable (1GB microSD card for less than 10 euros) and prices are dropping constantly. If someone decided to build a RAID card reader, we could easily get a foot in the door. For about 60 euros it would be possible to get something between a slowish but reliable 6GB flash drive or a speedy and snappy 1GB flash drive.

So why exactly didn't anyone thought of this? We already have IDE CF card readers, some models supporting 2 drives, that can be had for about 6 euros. Why not a RAID flash card reader?

Re:

[dfn_deux (535506)]

The difference is that not all NAND flash is created equally. The multi-layer cell type which is commonly used in commodity flash devices isn't nearly as fast nor as reliable as the single-layer cell type which is used in the highspeed drive replacements we are seeing hit the market now. The difference isn't trivial when an mtron SLC SSD can do about 5 times the throughput speed of competing higher density SSDs which use MLC nand flash.

I don't work for mtron, but I am a satisfied customer.

Re:

[takev (214836)]

It is called P2

http://en.wikipedia.org/wiki/P2_(storage_media) [wikipedia.org]

From the wiki: The P2 Card is essentially a RAID of SD memory cards

48GB Flash MCP?

[tgd (2822)]

I clearly am going to need to be set up for group EIGHT access. </flynn>

Not parallel

[RecessionCone (1062552)]

It's not clear if it's possible to write to them in parallel -- if so the device should be pretty damn fast.

It's pretty obvious that it's not possible to write to this array of chips in parallel, because you just can't fit enough pins in a tiny package to provide the necessary interface for talking to 24 chips simultaneously. Also, take a look at the picture from TFA: http://www.koreatimes.co.kr/upload/news/070905_p10 _hynix.jpg [koreatimes.co.kr] - you can see that all the leads to the different chips are wired to the same pads. This doesn't prove my point - they could all be power or ground connections, but looking at the comp

100,000 write cycles is plenty...

[tyme (6621)]

100,000 write cycles is plenty, so long as you buffer the writes and limit their frequency. All you need to do is either put a big honking RAM writeback cache next to the FlashRAM, or enforce writeback caching in the OS. If you can get the write frequency down to about two writes per hour, and do good load leveling, your FlashRAM will last for 5 years, which is about as good as most consumer grade hard disks (and possibly better, since the 'expired' FlashRAM drive could still be perfectly readable). Two wri

According to NASA

[brunes69 (86786)]

Hynix, has announced they have stacked 24 flash chips in a 1.4mm thick multi-chip package

According to NASA, it may even be possible to stack 48 chips in a 2.8mm package. Scientists also speculate someday we may be able to achieve up to 240 chips in a 14mm thick package.

Re:

[by Anonymous Coward]

The NAND gate is a union worker, also entitled to a smoke break every 3,000 write cycles.

Re:Why only 100,000 times

[jandrese (485)]

It's due to the way Flash works. A flash bit is basically a conductor surrounded by an insulator. To store a bit, you apply a large charge to the insulator to increase the charge of the conductor, basically your burning through the insulator to get your charge though. Once it is on there, to read the charge you have to apply another large charge to the insulator and see if the resultant charge is n or n + m. The m factor comes from latent charge on the conductor.

Anyway, the upshot of this is that because you have to constantly burn charge through the insulator to use the part, eventually you basically burn out the insulator and cause it to leak charge. Once it starts leaking, you lose your stored bits and the part is useless.

Re:

[treeves (963993)]

So is the summary (I didn't RTFA of course) misleading by continually referring to NAND flash having this limitation? Doesn't it also apply to NOR flash?
OK, I just looked at the Flash entry on wikipedia, and it appears that it's even worse for NOR flash.

Re:

[petermgreen (876956)]

Yeah memory that needs power to keep it's contents is known of as a ram. Much faster than flash and basically no limit on write frequenc but the boards to make it act like a disk are expensive and you have to make damn sure it keeps power if you have anything important stored there.

Re:

[Knara (9377)]

Keep in mind that the 100,000 writes number is a stock number that seems to go up all the time, not that this comment directly addresses the "why".

Re:

[Bacon Bits (926911)]

It's wholly one of mechanical endurance of the components, AFAIK. The gate is wedged, for lack of a better term. Everything physical wears out. It was much worse in the early 1990s, but whole orders of magnitude in improved performance have been made since then.

I've never seen a study conclude that the write limitation on NAND flash-based devices is a significant impact. Some of the studies have cited worst case scenarios of 50 years of continuous operation. It is far more likely that the device will p

Re:Why only 100,000 times

[networkBoy (774728)]

Flash is rated in erase cycles, not write cycles. Erase is the most damaging event to the tunnel oxide layer in the device, which is why they fail.
Flash Cell stackup (same for NOR and NAND, the interconnection of cells determines what type of array it is):

G - gate (metal)
ONO - Oxide/Nitride/Oxide layer
FG - Floating Gate (Poly)
tOx - Tunnel Oxide (very thin)
Si - wafer (NPN/PNP wells)

-nB

Re:NAND flash writes

[TinyManCan (580322)]

This is straight bollocks. Its ridiculous to think that you could only write to a NAND block 1000 times.

Commercial products in the high-end flash space are promising 500,000+ writes.

We are not talking about glorified thumb-drive flash memory here, but decent chips with good wear leveling and high quality construction.

Re:

[Surt (22457)]

http://download.micron.com/pdf/datasheets/flash/na nd/4gb_nand_m40a.pdf [micron.com]
Says 100,000 program/erase cycles right on the first page (though I do note they only 'guarantee' 1k writes for the first block).

Re:

[Bacon Bits (926911)]

You need to have either firmware on the device that handles wear leveling (this seems appropriate for IDE drives or as part of the spec for flash media standards) or use a file system which handles it for you.

One of the biggest offenders is file systems (such as the default configuration for NTFS) that track last access times. That information is all stored in the MFT for NTFS, so frequently accessed files will be writing to this table constantly.

Re:

[networkBoy (774728)]

Flash devices have a small microcontroller already embedded in them to control programming and erase voltages. It also is responsible for wear leveling (at least in NOR devices).

Re:

[TeknoHog (164938)]

Imagine a Beowulf cluster of 48 HP 48GXs with 48 GB each.