Intel and Micron Unveil 3D XPoint Memory, 1000x Speed and Endurance Over Flash 172
MojoKid writes: Today at a press conference in San Francisco, Intel and Micron unveiled 3D XPoint (Cross Point) memory technology, a non-volatile memory architecture they claim could change the landscape of consumer electronics and computer architectures for years to come. Intel and Micron say 3D XPoint memory is 1000 times faster than NAND, boasts 1000x the endurance of NAND, and offers 8-10 times the density of conventional memory. 3D XPoint isn't electron based, it's material based. The companies aren't diving into specifics yet surrounding the materials used in 3D XPoint, but the physics are fundamentally different than what we're used to. It's 3D stackable and its cross point connect structure allows for dense packing and individual access at the cell level from the top or bottom of a memory array. Better still, Intel alluded to 3D XPoint not being as cost-prohibitive as you might expect. Intel's Rob Crooke explained, "You could put the cost somewhere between NAND and DRAM." Products with the new memory are expected to arrive in 2016 and the joint venture is in production with wafers now.
Isn't electron based, it's material based? (Score:5, Funny)
And what material are they using? Positronic unobtanium?
Re: (Score:2)
Re: (Score:2)
Painted red.
Everyone knows red wunz go fasta!
Re: (Score:3)
No - green marker on the edge.
Re: (Score:2)
FLAMES! FLAMES on the SIDE plx.
memresistor? (Score:2, Interesting)
http://www.bbc.co.uk/news/tech... [bbc.co.uk]
"By contrast, 3D XPoint works by changing the properties of the material that makes up its memory cells to either having a high resistance to electricity to represent a one or a low resistance to represent a zero."
Sounds like a memresistor?
Seems Not (Score:5, Interesting)
from the The Register [theregister.co.uk]
Re:Seems Not (Score:5, Informative)
They can deny all they want - intel has been filing patents on a new type of GST PCM left and right over the last 6 years.
Re:Seems Not (Score:5, Insightful)
Nanoscale slider switches? ;-)
Seriously, though, it's some sort of material change according to what little information has been released.:
-- Tom's Hardware [tomshardware.com]
What's really interesting is the PDF with one diagram showing Xpoint sooner and then 3D XPoint on the 2018-2019 timeline at Semicon Taiwan [semicontaiwan.org] that later has a diagram much similar to Intel/Micron's diagram. It appears to be showing a variable resistor (potentiometer) then a diode between the word line and bit line crossbars.
If they are building a materials-based variable resistor that gets written to be more or less resistive based on voltage what are they calling that process? It needn't be chalcogenide, but it sure sounds like some other sort of phase change to change the resistance. If it is memory that adjusts its resistance based on past voltages and uses that resistance for reading the value, that sounds like a memristor. (According to Chua all PCM, ReRAM, and MRAM are memristors. [wikipedia.org])
I think perhaps Intel and Micron are saying it's not PCM and it's not memristors just so people don't confuse it with other attempts at similar but different approaches.
Re:memresistor? (Score:5, Interesting)
A non volatile PC would be nice.
No it wouldn't. Not unless we go back to having hard reset buttons on the front of our machines. The distinction between volatile and non-volatile memory is useful since we still have such shitty software full of bugs and security flaws. I wan't to be able to "reset" my machine without having to erase my hard disk.
Re: (Score:3, Informative)
Additionally with a non-volatile memory the system could be "suspended to disk" / hibernated simply by syncing all pending writes and powering off. In most cases a 1000 ms operation. Resuming would have similar performance. The machi
Re: (Score:2)
In most cases a 1000 ms operation.
That seems terrible that we have to wait for a second, even assuming we're talking about NAND rather than RAM-speed here.
Re: (Score:2)
Why not try to solve those problems _too_?
Re: (Score:2)
I'd say it has worked, at least better, for, oh, the most valuable company on the planet.
Memory paritions. (Score:2)
Re: (Score:2)
Newegg (Score:5, Informative)
As with any new 'pewter tech, I'll believe it when when it I see it on Newegg with >500 reviews, > 3.5 stars, and affordable for the average Jane/Joe.
Re:Newegg (Score:5, Interesting)
Usually i'd agree... there's been countless up and coming new types of memory that never make it.
But i'm cautiously optomistic here because
a) It's Intel and not some tiny obscure VC
b) they said they already have wafers and mention 2016 O_o !
no wonder they ditched their awesome SSD controllers.
Re: (Score:2)
Usually i'd agree... there's been countless up and coming new types of memory that never make it.
But i'm cautiously optomistic here because
a) It's Intel and not some tiny obscure VC
b) they said they already have wafers and mention 2016 O_o !
no wonder they ditched their awesome SSD controllers.
It's DRAM that's in the crosshairs.
Re: (Score:2, Insightful)
> It's DRAM that's in the crosshairs.
Only to a small extent. This would reduce the need for DRAM cache of SSD data. Computers will still need huge amounts of DRAM for workspace. Workspace memory needs trillions of times more write cycles than this provides.
Re: (Score:2)
> It's DRAM that's in the crosshairs.
Only to a small extent. This would reduce the need for DRAM cache of SSD data. Computers will still need huge amounts of DRAM for workspace. Workspace memory needs trillions of times more write cycles than this provides.
Or more SRAM cache local to the CPU with cache lines being merrily lobbed twixt the SRAM and the magic new memory. Maybe. A non volatile PC would be neat.
Re:Newegg (Score:4, Insightful)
It honestly depends on how they measure endurance. If it's measured as 1000x the 3 million writes, then no. If it's 1000x the three-year estimated wear-out time under consumer conditions, then that's phenomenal.
Re: (Score:2)
It honestly depends on how they measure endurance. If it's measured as 1000x the 3 million writes, then no. If it's 1000x the three-year estimated wear-out time under consumer conditions, then that's phenomenal.
It doesn't make sense to quote endurance form an consumers perspective when it's not a product yet (as either write cycle or life span) - that's all highly dependent on what the size of the device is, how well the controller is designed and how much it utilises high endurance caches to reduce wear, NAND or not - it still has an endurance where these factors will play a role.
Worst possible case TLC NAND is 1,000 cycles on the page level, that would make this memory 1,000,000 at worst which is at least 10 tim
Re: (Score:2)
Re: (Score:2)
Re: (Score:3)
As with any new 'pewter tech,
this one is also based on alloyed tin.
A memristor or something else? (Score:3)
Looks Great, Beware? (Score:2)
Re:Looks Great, Beware? (Score:5, Funny)
Of course it's a government conspiracy.
Who do you think they orchestrated the phase out of tin foil and replaced it with alunimium foil?
Soooo you're saying (Score:2)
SSD's are about to go to ludicrous speed.
Re:Soooo you're saying (Score:4, Funny)
Wake me when they've actually gone plaid.
Re: (Score:2)
This is it. They're using plaid technology. SSDs were already at ludicrous speed.
Is it the start of the Diamond Age? (Score:2)
.
bottlenecks (Score:3, Informative)
Re: (Score:3)
From the BBC:
Rather than pitch it as a replacement for either flash storage or RAM (random access memory), the company suggests it will be used alongside them to hold certain data "closer" to a processor so that it can be accessed more quickly than before.
Re:bottlenecks (Score:5, Informative)
M.2 is the desktop interface for this, it supports 4 PCI 3.0 lanes at 985 MB/s per lane that is nearly 4GB/s. PCI 4.0 is not to far off and doubles that.
Re: (Score:2, Informative)
http://www.overclock.net/t/1489684/ssd-interface-comparison-pci-express-vs-sata
It's almost like that problem has already been thought of. Go figure.
Re: (Score:2)
http://www.overclock.net/t/1489684/ssd-interface-comparison-pci-express-vs-sata
It's almost like that problem has already been thought of. Go figure.
Plus new Intel chips (I think) have more PCI-e lanes. I'm hoping that rumor is true, anyway.
Re: (Score:2)
Re: (Score:2)
You just attach it directly to the PCIe bus with a dedicated controller like PCIe SSDs today rather than through the existing SATA ports http://www.engadget.com/2015/0... [engadget.com] . Over time, the other interfaces will catch up. x4 Gen3 PCIe goes pretty quickly...
Re:bottlenecks (Score:5, Informative)
SATA 3.2 (aka SATA Express) is a connector change, but is actually PCIe. PCIe is already fast enough. IIRC, Apple hooks up some SSDs directly through PCIe.
And, PCIe can actually go "off board" via a cable (since PCIe is based on separate upstream/downstream lanes and differential line drivers). Also, PCIe 4.0 will have a transfer rate of 31.5 GB/s, yet be fully backward/forward compatible.
Intel already has a CPU package that has two substrates wire bonded together, one for CPU and one for memory. When I saw this, I assumed it would be to accomodate HP's memrister memory. But, now, it's [obviously] been planned for this new type of memory.
Re: (Score:2)
One of the articles says the initial products will be PCIe and NVMe.
The Toms Hardware Article is much better:
http://www.tomshardware.com/ne... [tomshardware.com]
Intel indicated the new memory would connect to the host system via the PCIe bus, which is yet another reason that Intel and Micron have been vocal proponents of NVMe. The NVMe protocol was designed from the ground up for non-volatile memory technologies, and not NAND in particular. Now it is apparent that Intel and Micron were laying the groundwork for something more as they developed the new protocol.
Clearly this memory will necessitate new motherboards. But I would also love to see this on Nvidia cards.
Re: (Score:2)
Re: (Score:2)
put me name down! (Score:2)
I'll have some o' that!
How does it compare to RAM, though? (Score:4, Interesting)
Intel and Micron say 3D XPoint memory is 1000 times faster than NAND, boasts 1000x the endurance of NAND, and offers 8-10 times the density of conventional memory.
How does the speed compare to conventional memory, though?
Re: (Score:2)
How does the speed compare to conventional memory, though?
Well, SSD typically have speeds in the range of 100s of MB/s and graphics cards in 100s of GB/s so if it's 1000 times faster than NAND I'd say pretty close.
Re: (Score:2)
Is it going to matter much? (Score:2)
We've already got non-volatile memory with extremely high endurance on the mass market (SLC NAND), so what you basically get out of this stuff is "It's like flash, but much faster."
The question becomes, what is enabled by having much faster flash memory? Sure, you might see some minor power efficiency increases in mobile devices if you don't need to keep the RAM powered, but that's not exactly world changing.
I'm not saying this isn't good, just that people are hyping it up, and I'm trying to determine what
Re: (Score:2)
Good point. When they say this is faster and has better endurance than NAND, did they mean TLC, MLC, or SLC? You are right that SLC NAND has good endurance and speed, but maybe this is better? The ultimate memory would be memory that is fast enough to eliminate the separate concepts of "RAM" and "storage" entirely, and make it one thing. Then instead of having 8GB of RAM and 2TB of storage, you would have 2TB of "storage RAM." This might be fast enough and reliable enough for that. Or maybe it is clos
Re: (Score:2)
Even if it's a thousand times more durable than NAND it's not much in a loop, if you just write to the same memory location over and over with DDR4 you can write every 5 cycles @ 1.25ns/cycle = 160,000,000 writes/second. I would think the greatest advantage would be a write cache which could return ~1000 times faster from a flush() making sure it's committed to non-volatile memory. The SSD can then work "behind the scenes" to move it to slower SLC/MLC/TLC.
Re: (Score:2)
But at that point you're basically just replacing the current approach of a supercapacitor and DRAM with some of this new stuff. You might save a few bucks, but that's a relatively small difference.
Re: (Score:2)
As long as your super cap is large enough to power the device while the entire DRAM cache is written to NAND (or your DRAM cache is small enough)
That could also mean performing erases as well as additional writes, to fit new data where old data used to be.
Or you could reduce your DRAM cache significantly or even remove it entirely and add some of this new memory to get faster sustained sequential write and less wear on the NAND.
Re: (Score:2)
Even if it's a thousand times more durable than NAND it's not much in a loop
That's why I said to use the DRAM as a cache. But maybe it just isn't good enough even with that.
Re: (Score:2)
Would that really make much of a difference? Computers already act, for the most part, like they have non-volatile memory. When you shut the lid on a laptop, it writes RAM to disk and goes to sleep. If you wake it up without having cut the power, it wakes up quickly. If you pull the power/battery, it takes a few seconds longer. In either case, it wakes up where it left off.
There is also nothing stopping developers from doing what you describe right now. Storage is fast enough that changes to most files can
Re: (Score:2)
Storage is fast enough that changes to most files can be saved directly to disk as they're made. When working in the cloud, this sort of "every keystroke saved" thing is already the norm.
Not quite: Just open a 500MB word document and insert a single character at the top of the file. Google Docs can't even handle this kind of file - delays are seconds to minutes. Even if it could work with a word processor, it won't work with everything. When I play MineCraft, it takes 30 seconds to save my changes. You can't commit my actions to disk at 60fps. Same thing with Photoshop or SolidWorks.
However, you made me realize that in-memory structures are vastly different from structures on disk to a
Re: (Score:2)
> Not quite: Just open a 500MB word document and insert a single character at the top of the file.
COW filesystems would have no problem with that scenario, especially when they have dynamic block sizes. There might still be some nasty write amplification (such as writing kilobytes of data to insert the one character), but it wouldn't be any slower than appending one byte to the end of the file.
Re: (Score:2)
SSD's are fast because they have 8 or 16 chips in parallel.
If you want to cram that in to a smaller form factor, you can't.
If you can figure out an interface, you could have a 128GB SSD in the form factor of a micro SD card that performs better than current NAND based SSD's
Re: (Score:2)
Re: (Score:2)
It sure sounds like the outcome could be cheaper, faster, more reliable and possibly even denser storage. How about a 10 TB drive that can saturate a SAS link for the price of a consumer 1 TB SSD now? It sounds appealing to have 40 TB of home storage at performance levels that would make a $200k enterprise storage buyer jealous.
Or that makes for 240 TB enterprise san shelf for the price of an existing 10 TB flash/rust hybrid shelf at speeds that will melt 16 gig fiber channel?
And who knows what value fast
Re: (Score:2)
Why would you expect it to be cheaper or denser than NAND? IMFT says it'll be more expensive than NAND, and even if the cost drops over time, so too will the cost of NAND.
Re: (Score:2)
I might expect some cost reductions because the increased durability will lessen the amount of excess memories needed for remapping when cells go bad. And don't larger drives use NAND chips in parallel for speed? If you can simplify packaging by using a single chip you might cut costs there, too.
If its as supercalifragilisticexpialidocious as they say it is, you might also expect enterprise adoption to increase, lowering the cost of NAND by cutting demand or resulting in more reliable NAND.
It's also hard
Re: (Score:2)
If I can buy a TB of RAM that's maybe DDR4 speeds but not DDR5 speeds but at say somewhere between NAND and DRAM pricing that would be huge.
Also incredibly useful for something like a phone where you might want to shoot 4k video. The CPU would have a hard time processing that but if you buffered to say a 64GB cache and then processed you could shoot highspeed for a minute instead of 2-3 seconds.
Re: (Score:2)
Why would most people need a terabyte of RAM? There is such a thing as diminishing returns, and you have to be doing pretty crazy stuff to need more than 32GB or so. Ultimately making RAM cheaper might save a few bucks, but that's not justifying the massive hype that non-volatile memory has had over the years.
In terms of 4K recording on a phone, video encoding is done in hardware, not software, and no amount of cache is going to solve that. If you're talking about highspeed, your 64GB cache would store arou
Re: (Score:2)
I'm rendering a project right now that has over a hundred instanced trees in a forest. So the forest is pretty much instanced, but each tree instance is around 1GB of memory and there are about 12 individual models. Then once I get into terrain geometry and villages and trains and that isn't even getting into the volumetric sparse oct-trees for like smoke from chimneys... etc.. anyway long story short 32GB is already gone *with* massive scale instancing.
Re: (Score:2)
Oops I mean "over a hundred million instanced trees" obviously haha.
Re: (Score:2)
You're far from a typical user, however, and 100 petabytes of this new stuff would be just as cost prohibitive as 100 petabytes of NAND or DRAM...
Re: (Score:2)
This memory is byte addressable (e.g. RAM-->random access memory), so no "block erasure" needed in the write cycle as in NAND flash. It's also 1000x faster than NAND flash (at 2ms), so access time should be about 2us, and no wear leveling needed. It also has a higher memory density--9x if you believe the block diagram. It can also be stacked 3D, which, IIRC, flash can't [or hasn't been] up to this point.
There are a number of other non-volatile "solid state" memory technologies in the works: magneto-r
Re: (Score:2)
Samsung started shipping 3D NAND in consumer SSDs in mid-2014 with the 850 Pro. It used a 32-layer 40nm process. It was their second-gen 3D NAND, with their first-gen 24-layer product released in 2013 for enterprise SSDs.
IMFT and Toshiba are both sampling 3D NAND, 32-layer for IMFT and 48-layer for Toshiba. Neither is saying what size, but some googling indicates the consensus is somewhere between 35nm and 50nm.
You're still going to need to do wear leveling. If you can write a byte every 2us, but you only
Re: (Score:2)
NAND will eventually hit the "die shrink" wall. Since this new Intel memory apparently fits nine cells in the same die area as a NAND cell, it will eventually take over from NAND.
As a side note [to show I'm not totally against NAND flash], a Japanese researcher found, about a year ago, that if you add a heating element to a NAND cell [similar to the one in ferroelectric memory], you can "boil off" the excess trapped charge and eliminate the "wear out". He believed that this was a trivial addition to exist
But is is crystalline? (Score:3)
But is is crystalline and available in pretty glowy colours? Because otherwise I don't want it. For too long we have been waiting for our 3D crystals that are memory devices in science fiction, and now we better get 'em!
Retention time? (Score:4, Interesting)
How long is it able to retain the data and under what range of conditions? Currently this is one of the big problems with flash, where small-process TLC memory is so fragile that reading it damages the contents, much like core.
Yay! (Score:2)
Finally, a solid replacement for NAND: I was past the point of despair.
The patent points to PCM (Score:2)
The patent [google.com] points to PCM. Maybe they're applying the same X-Point structure to a different material system, though. I'm guessing, however, that they're just saying it's not PCM in an attempt to dodge other patents.
Thank me (Score:5, Funny)
You can thank me for this being announced now, because I *just* bought a SSD last week. You're very welcome.
Likelyhood of ACTUALLY seeing it in stores? (Score:2)
Even if it's intel, rather than a tinpot outfit, I think it's very unlikely we're going to see this product on shelves ready to use in less than 4 to 6 years.
What's it for also? It's going to wear out? So using it as standard RAM is probably not practical. Should it simply be a replacement to SSD's? an even faster primary storage device?
Re:Ohh Ohh (Score:5, Funny)
And I already know how we're going to erase [pinimg.com] that memory.
Re: (Score:2)
Re: (Score:2)
Re: (Score:3)
I don't usually bother to make negative comments, but this is too much. If their claim is true, then the memory would cost the same per bit of memory and 8 to 10 times as much per unit of volume. Got it? And it is not really important how much it costs per unit volume. It's just important that the memory will not take up much space. You really need to slow down and work on your critical thinking skills. /rant
Re: (Score:2, Informative)
Re: Cost? (Score:2)
and $3.30 is fine. I paid way more for my SLC SSD's for my primary ZFS log device. Which are still going strong - thank you Intel. Gimme a giant one of these to swap out the mostly unused VM memory pages that are lost to containerization. It's webscale RAM. :P
MRAM? (Score:5, Interesting)
This '3D Xpoint memory' sounds very much like MRAM as described by the following article
http://spectrum.ieee.org/semic... [ieee.org]
Last year (2014) Samsung reportedly was collaborating with 15 partners in developing similar spintronic MRAM memory technology
http://www.mram-info.com/samsu... [mram-info.com]
Hynix and Toshiba also partnered to develop their own version of MRAM
http://phys.org/news/2014-04-f... [phys.org]
In less than 5 years we might get to enjoy the fruits of the labor of the thousands of researchers who have been working very hard to make the spintronic dream come true, and I for one, wish to take this chance to thank them for their hard works!
Ram replacement? (Score:5, Interesting)
There has been some discussion http://bit.ly/1SLtYAh [bit.ly] that 3D XPoint might be a replacement for RAM in mobile devices because it is "ram-like" in speed and low power due to the fact that it's non-volatile. If this can replace RAM and NAND in phones and tablets, it will be a major milestone in the history of computing.
Re: (Score:3, Insightful)
At "up to" 1000x the endurance of NAND? Who wants RAM cells that die after "up to" 300k writes?
Re:Ram replacement? (Score:4, Informative)
If we are talking MLC NAND, you are off by a factor of 10. http://www.anandtech.com/show/... [anandtech.com]
But since the XPoint cells are individually addressable bit by bit, the comparison is probably to SLC, in which case the relevant number of P/E cycles for XPoint would be 10E8 and you are off by a factor of more than 33,000.
Re:Ram replacement? (Score:5, Insightful)
1 Performance difference based on comparison between 3D XPoint technology and other industry NAND
2 Density difference based on comparison between 3D XPoint technology and other industry DRAM
3 Endurance difference based on comparison between 3D XPoint technology and other industry NAND
They don't actually specify *what* NAND.
So assuming typical marketing weasels, that means it's 1000x the worst TLC they could find anywhere - and that's typically rated at 300 P/E cycles.
But since the XPoint cells are individually addressable bit by bit, the comparison is probably to SLC
Yes, they totally wouldn't write "1000x the endurance of industry-leading IMFT SLC NAND" if that were the case...
in which case the relevant number of P/E cycles for XPoint would be 10E8 and you are off by a factor of more than 33,000.
Would you like a bridge with that?
Re: (Score:3)
Hmmm. You do have a point. :)
Re: (Score:2)
Don't forget that when new technology is announced they always list the upper limits of a technology. So it has 1,000 times the potential of current best-case-scenario NAND but you won't see that 1,000 performance boost for 3 decades when they tap out the technology's maximum potential.
Re: (Score:2)
I think you mean 333 in the latter case?
Re: (Score:2)
Yes, you are correct. But now I have seen that Micron says the endurance is not related to write cycles. This gets better and better.
Re: (Score:3)
The 1000x durability wasn't based on writes alone. In fact, they even say it's "not significantly impacted by the number of write cycles"
Re: (Score:2)
Also note that the cost per bit for XPoint is lower than for existing DRAM. If the endurance is not cycle related and the speed is close and the power use is lower, this would be a clear improvement.
Re: Ram replacement? (Score:2)
Who cares? Your carrier will still stop shipping security updates before your RAM error map is full.
Re: (Score:2)
Material based makes it sound like the Ovonics (Phase-change [wikipedia.org]) memory that has been bandied about since like forever by Intel.
The performance seems to be more in line with MRAM though. I suspect some of those claims they make are bullshit. But we'll see.
Re: (Score:2)
This could be it: Resistive random-access memory [wikipedia.org] i.e. ReRAM.
Quoting Wikipedia:
"The basic idea is that a dielectric, which is normally insulating, can be made to conduct through a filament or conduction path formed after application of a sufficiently high voltage. The conduction path can arise from different mechanisms, including vacancy or metal defect migration. Once the filament is formed, it may be reset (broken, resulting in high resistance) or set (re-formed, resulting in lower resistance) by another v
Re:cost per bit... (Score:5, Interesting)
I can see this being used two ways:
A fast SSD.
A swap device/slow RAM.
This can make things interesting for SANs, especially because it adds another tier to the disk type hierarchy.
I'd like to see it used as a cache, as well for swap and the core OS files so booting is made quicker. However, it would be useful for database index volumes as well.
Re: (Score:2)
PalmOS (Score:5, Informative)
However, current operating systems and programming techniques aren't up to this yet. It will take a long time.
PalmOS has been 100% RAM-only from the original Palm Pilot all the way up to Palm Thungsten III (Palm T5 with Flash, and Palm Live with a micro drive where the first to actually have a permanent main storage).
Everything is in-RAM, everything is stored in in-RAM databases. Data saving is immediate, etc.
(Also, although not so extreme:
lots of embed system, usually Linux-based, only have a minimal amount of ROM as sole storage and mainly work using RAM. Though they aren't completely in-RAM oriented and still use the concept of "files" and "storage", and thus make use of ramdisk (usually tmpfs) to hold files.
Still, that also machine which mainly count on RAM storage).
Re: (Score:2, Insightful)
Too good to be true? The guy's name is Crooke...
Well, we don’t hold it against you that your name is Wanker.