Tuesday, March 28, 2017

What is Intel's Optane memory?

Intel has developed a new product, dubbed Optane, that is a memory technology that supposedly combined the speed of conventional DRAM with the nonvolatility of flash memory.   It would appear that the devices function as a form of "3d crosspoint" memory, where the functionality is all in a blob of material at the crossing point between two wires (a bit line and a word line).  Depending on some particular voltage pulse applied to the junction, the blob of material either has a high electrical resistance or a low electrical resistance, corresponding to the two different states of a binary bit.   The upsides here are that information is stored in a material property rather than as charge (making it non-volatile, probably radiation hard, probably uninfluenced by magnetic fields, etc.), and the devices can be packed very densely, since they need fewer transistors, etc. than conventional DRAM.

There are multiple different mechanisms to achieve that kind of response from the mysterious blob of material.  For example, you can have a material that changes structural phase under current flow, between two different structures with different electrical properties.  You can have a material where some redox chemistry takes place, switching on or off a conductive filament.  You can have a material where other redox chemistry takes place, along with the migration of oxygen vacancies, to create or destroy a conductive filament (as in the HP implementation of memristors, which I've written about before).  You could use magnetic data storage of some sort, with spin transfer torque driving switching of some giant magnetoresistive or tunneling magnetoresistive device.

Breathless articles like this one this week make some pretty bold claims for Optane's performance.  However, no one seems to know what it is.  There has been speculation.  Intel's CEO says it's based on actual bulk changes in the electrical properties of the mysterious material.  Intel categorically denies that it is based on phase changes, "memristor" approaches, or spin transfer torque.  

Well, now that they are actually shipping chips, it's only a very short matter of time before someone cuts one open and reverse-engineers what is actually in there.  So, we have only a little while to speculate wildly or place bets.  Please go for it in the comments, or chime in if you have an informed perspective!  Personally, I suspect it really is some form of bias-driven chemical alteration of material, whether this is called "memristor" in the HP sense of the word or not.  (Note that something rather analogous happened back when IBM and Intel switched to using "high-k" dielectrics in transistors.  They wouldn't say what material they'd come up with, and in the end it turned out to be (most commonly) hafnium oxynitrides.)


Anonymous said...

In case you haven't seen this, Doug!


Yup, I know :)

Douglas Natelson said...

Thanks! This one is also relevant, from the Micron collaborators: http://www.flashmemorysummit.com/English/Collaterals/Proceedings/2011/20110811_S303_Atwood.pdf

So, the magic words seem to be "ovonic threshold switching in chalcogenides". Here is a review paper with Mott (!) as a co-author: https://journals.aps.org/rmp/abstract/10.1103/RevModPhys.50.209
I guess they can split hairs about whether this is really phase change memory if the structure of the switchable material is amorphous in both the on and off states.

Anonymous said...

Thanks for the paper link!

The state changes can occur differently in the selector vs. the memory cell - with an electronic mechanism in the selector but with a physical state transition in the memory cell, the latter by virtue of Joule heating. I guess that gives the marketing license to call it 3D Xpoint instead of Phase Change, but its probably poor over-selling/hype of the Phase Change moniker from the past - leading to the different marketing name.

Talking of hype (BS, in fact, IMHO) though, the there is probably no bigger one than renaming resistive memory/switch as "memristor", pretending that to be some Nobel-worthy "discovery", while resulting in countless Nature papers. Relative to that, you'd have to give Intel/Micron credit for delivering a solid-state product (pun intended) after long but fruitful R&D.