Slashdot Mirror
Why Micron/Intel's New Cross Point Memory Could Virtually Last Forever
Lucas123 writes: As they announced their new 3D XPoint memory this week, Micron and Intel talked a lot about its performance being 1,000X that of NAND flash, but what they talked less about was how it also has the potential to have 1,000X the endurance of today's most popular non-volatile memories. NAND flash typically can sustain from 3,000 to 10,000 erase-write cycles — more with wear-leveling and ECC. If Micron and Intel's numbers are to be believed, 3D XPoint could exceed one million write cycles. The reason for that endurance involves the material used to create the XPoint architecture, which neither company will disclose. Unlike NAND flash, cross point resistive memory does not use charge trap technology that wears silicon oxide over time or a typical resistive memory filamentary architecture, which creates a statistical variation in how the filaments form each time you program them; that can slow ReRAM's performance and make it harder to scale. Russ Meyer, Micron's director of process integration, said 3D XPoint's architecture doesn't store electrons or use filaments. "The memory element itself is simply moving between two different resistance states," which means there's virtually no wear.
I believe they've also indicated the "1000x endurance" figure isn't based upon write cycles -- meaning it has to be based on typical lifetimes. So yes, you're talking many years of service.
Kythe
"Crappy power" is normal. Manufacturers need to design for that. How long something lasts in a lab is irrelevant.
that there is always a catch.
So, what's the catch?
At the moment - availability.
Not yet determined - cost
"She's furniture with a pulse"
Even at 1000x faster, i am not sure its fast enough either. But would we ever want persistent ram for everything? I mean how could we turn it off and turn it on again to fix anything ;)
If information wants to be free, why does my internet connection cost so much?
If information wants to be free, why does my internet connection cost so much?
Your Internet connection is not information. It is a complex system of wires/tubes/fibers run by computers, and uses electricity, occupies land. It is operated by a corporation who pays people, negotiates with other corporations, and deals with/pays for many subtle and not so subtle political aspects of the whole thing.
This issue is a bit more complicated than you think.
But a small boot routine in ROM that erases a range of RAM (persistent or not) isn't that hard to conceptualize. Besides, depending on the type of volatile RAM, it doesn't always come up as all zeros at power-on either... I mean, what do you think happens when you press a reset button? Everything is still in RAM at that point.
What would make things different is a software architecture change that gets rid of the separate permanent storage layer and makes everything RAM-persistent. That would be kinda strange to imagine.
I just want an SSD made with this ASAP. I hope they don't start out at $500 for a 100GB drive.
A brain is a terrible thing to waste... Mind? That's debatable.
"Crappy power" is normal. Manufacturers need to design for that.
Yeah, see physics doesn't care what you think is "normal". While it's possible to design for a reasonable range of power conditions it is not economically possible to design for all of them. Frankly if your power quality is so poor that they are constantly blowing out light bulbs the answer is to fix the power, not the bulbs. You take the bullets out of the gun rather than insist everyone wear a bullet proof vest.
Is this memory based on silicon, or something else, like GaAs or Germanium or Graphene or something else?
Given that they've released close to zero technical details on how it works, but stated that it's nonvolatile, has 1000x the endurance of NAND flash while being 1000x faster, is cheaper than DRAM, and will be available in 128GBit capacities any minute now, my guess is that it's based on magic.
I have not done the sums, but I doubt that 1000x NAND's endurance is sufficient for also replacing DRAM.
No need to do any "math". DRAM is written millions of times per second, this would only last a few seconds as a PC's main data store.
No sig today...
""Crappy power" is normal. Manufacturers need to design for that. "
Get a UPS.
It's going to cost more than NAND flash.
But it would make a GREAT cache for spinning rust. None of the longevity problems of NAND, 1,000 times faster. Ka-chow.
Faster than *THE* FLash? Doubtful. Or did you mean Adobe Flash? That's not much of an achievement.
Defending IP by destroying access to it? That makes sense, RIAA/MPAA. Go to the corner until you can play nice!
Can someone explain how exactly they're "moving between two different resistance states"? Because I think that that in itself does not guarantee lower wear.
Has anyone heard anything technical about how this works?
Given that they've released close to zero technical details on how it works, but stated that it's nonvolatile, has 1000x the endurance of NAND flash while being 1000x faster, is cheaper than DRAM, and will be available in 128GBit capacities any minute now, my guess is that it's based on magic.
Until they release full specs you cannot assume that it's based on magic. I just hope they didn't base it on myth. But we'll see.
It's going to cost more than NAND flash.
But it would make a GREAT cache for spinning rust. None of the longevity problems of NAND, 1,000 times faster. Ka-chow.
For that matter, it would be a pretty good cache for NAND SSDs. I could do with most of my writes being 1000X faster.
Note to ACs: I usually delete AC replies without reading them. If you want to talk to me, log in.
Isn't that because of the memory refresh, which wouldn't be needed for that kind of memory?
"we are all atheists about most of the gods that societies have ever believed in. Some of us just go one god further."
Practically unlimited? 1000x the endurance of NAND gets you 3 million writes. It's per-byte addressable, and it's supposed to have performance 1000x that of NAND. If you assume NAND has typical access times of 2ms, and then this stuff could write a single byte up to 500,000 times per second.
In short: without wear levelling, you can burn out this stuff in 6 seconds if you overwrite the same byte as fast as possible.
If you implement wear levelling on a block level, and use it for storage (and not to replace RAM), then it should be effectively unlimited write cycles.
Here something I learned a while ago; Speed isn't how fast you do something (it is, but only partially), it is often a measure of whether or not you actually CAN do something.
Here is my story:
In the Mid 90's I ran an ISP. Part of my daily chores was processing logs looking for anomalies, and to gather stats needed to project out the upgrades that are needed. When I started, the logs were small and it took a few minutes to process. As the business grew, the process took longer and longer. It soon took hours to process the logs for the day. It became so problematic, that I just stopped doing them.
But business kept growing, and I needed the stats. So, I bought a new machine. The new machine could process the logs in five minutes, what took hours on the older machine. Mind you, this was one generation difference between the two machines (68040 to PPC 701), but that was all that was needed to show me that speed isn't just how long it takes, sometimes it is whether or not you do the thing you ought to do.
Seeing the price of SSDs and Spinning HD, at their current price points, there is no reason to NOT get the SSD, at whatever cost they are now. Especially for enterprise grade systems that need the IOPS, Even at $1000 for 1 TB SSD is extremely affordable speed, especially when considering you get 90,000 IOPS.
IF we're talking about 1000x faster, the speed is enough to change what we can do.
Agent K: A *person* is smart. People are dumb, stupid, panicky animals, and you know it.
I've been saying this for a long time. There is a definitive hierarchy between all the different memory locations. Unfortunately we don't have an OS that looks at all these levels as one. We have abstarcted all the CPU Cache, RAM, NAND, Spinning disk, clout etc as separate levels, rather than a single level with varying degrees of capability.
When we have an OS that can view all the levels as one, intelligently, we'll have a much more efficient OS. It might take a whole new design from the hardware up to accomplish.
Agent K: A *person* is smart. People are dumb, stupid, panicky animals, and you know it.
This is a good point. If it is significantly more expensive, say 1GB of this as a write buffer on a large SSD drive will make the NAND drive last nearly forever, as frequent writes can be buffered and only written to the nand when necessary. The biggest issue with NAND is when software constantly writes to the disk, and pushes the write wearing logic to the limit.
Unofficial sources are starting to say that XPoint does not exhibit wear from write cycles and the "1000x more endurance" is normalized to some other metric. If it was normalized against time, then you may expect NAND to last 3-5 years which would put XPoint around 3,000-5,000 years life time. We won't know until more official data or hands-on reviews happen.
If latency is 1,000 times lower and endurance is 1,000 times higher then, under continuous load, endurance measured in real time is unchanged. Not by any means a hypothetical scenario.
When all you have is a hammer, every problem starts to look like a thumb.