Back in June, I predicted that Intel (NASDAQ:INTC) and Micron (NASDAQ:MU) were close to releasing a new, nonvolatile memory technology that would disrupt the fundamentals of computing price, performance and power efficiency. And I was right: the duo announced 3D XPoint ("3XP" for the sake of brevity) just over a month later. But, during the sloppy announcement, they categorically denied that it leveraged the phase change memory ("PCM") technology that I had predicted.
Based on the substantial number of patent applications and research papers, I will stick with my prediction that 3XP is just PCM under a technicality that they are leveraging in order to distance themselves from the intellectual property theft that I outlined in my June article (again, I'm not saying that Intel or Micron were behind that theft, but they did become the beneficiaries).
That article is necessarily a prerequisite for this one in order to bring yourself up-to-speed on the origin of PCM at Energy Conversion Devices (ECD). If you don't want to read the whole thing, just know that PCM has been developed by ECD over the last 50 years, reaching commercialization through their Ovonyx subsidiary, which has since been acquired by Micron.
Never Mind the Man Behind the Curtain
During the 3XP announcement, there was indication that 3XP was not phase change memory because it leveraged some other technology described as a "bulk material property change". To understand this better, we need to take a look at the previous state of the art in PCM: the "mushroom cell".
Take a look at Figure 1 in that last link. Figure 1(a) is an actual electron microscope picture of a phase change memory mushroom-type cell. Figure 1(b) and 1(c) are schematics of the important portions of the mushroom cell where the yellow/brown is the phase change material and the electrodes are grey. Notable here is that the only material that changes phase is the mushroom top nearest the lower electrode (the brown portion). Figure 1(d) is what a hypothetical "bulk" cell would change.
In order to qualify as a "bulk material property change", the whole set of phase change material would need to change - not just the mushroom-shaped subset. Micron's only other viable near-term memory technology is CBRAM (which is unfortunately also known as "Programmable Metallization Memory" - abbreviated PMC, which is decidedly unhelpful in mixed discussion with PCM).
But we know from Micron's 2015 Winter Analyst presentation that CBRAM is scheduled for 2017. How do we know this? If you'll turn to slide 18, you can see that they have "New Memory A" scheduled for 2015 while "New Memory B" is scheduled for 2017.
We know now that 3XP is the former. Turning again to slide 65, we can see that the volatility of "Performance Focused" new memory is closer to that of DRAM than to flash so it is safe to say that 3XP is not this technology.
So the "Performance Focused" is the one scheduled for 2017. In perusing Micron patents, you'll see that CBRAM is the only one with high volatility (it actually has variable volatility, as outlined in section  in this Micron patent - note that, while CBRAM is the exclusive technology listed in claim 8, Micron uses the term "resistance change memory" thereinafter in order to make the patent generically applicable to other resistance-based technologies).
While there is a form of bulk-switching ReRAM known as correlated electron RAM, this doesn't fulfill the Intel/Micron "exclusive" attribute that they have advertised. And it does not have high volatility.
A Rose By Any Other Name
Because CBRAM is not a bulk process, I'm revisiting PCM. The underlying material in PCM is chalcogenide (warning: audio) - this is the stuff that changes its phase from a low-resistance "crystalline" phase to a high-resistance "amorphous" phase. This material is also used in rewritable optical (CD, DVD, BluRay, etc) media, leveraging a change in reflectivity between these phases - fascinating stuff!
After more than 50 years of intensive research, phase change science still has more questions than answers - we're still learning the mechanisms by which chalcogenide performs. With traditional PCM, the phases are changed by way of "nucleation": joule heating is manipulated to expose the memory cell material to the different temperature conditions that are required to change from crystalline phase to the amorphous phase (and also to intermediate states that are required for storing multiple bits per cell). But nucleation through joule heating is energy-intensive and slow.
But, if you dig just a little bit, you'll discover that new bulk phase change mechanisms have been discovered:
Switching mechanism in amorphous chalcogenides
M Popescu and I D Şimăndan
© 2010 IOP Publishing Ltd • Journal of Physics: Conference Series, Volume 253, Number 1
The electrical switching in solid compositions based on chalcogenide materials is discussed. Different mechanisms proposed for switching effect are reviewed. A new description of the switching phenomenon is done. The switching is regarded as due to formation and breaking of the links between the dendrites of crystalline nuclei in bulk materials, as a consequence of the energy pumped by an electrical field. This mechanism explains the very short switching time (<20ns), the possibility to get smart memories based on multisteps of resistivity and the high number of cycles supported by the cell (1016).
This is a very technical report and it is a good read for anyone with a proper background - CBRAM metallization is outlined and a 16-level (4 bits per cell, QLC) PCM cell is demonstrated. But, to the point, it clearly outlines a new bulk switching mechanism for PCM: growing nano-crystals electronically (i.e. - using electrons instead of heat). This newer switching mechanism is well-known as "nano-phase change" in the PCM research community.
This is just phase change on a much smaller scale. They first discovered this phase change mechanism in 1962 so it isn't new but it does provide a technicality that allows Intel and Micron to deny that 3XP is traditional PCM. Regardless, the benefits of nano-PCM are manifold: lower power, faster, more durability and longer retention than the previous "mushroom-cell", filamentary PCM. This is reinforced in a recent Ovonyx patent application (note again that Ovonyx is Micron-owned but this patent has at least one Intel inventor listed):
- Also, in some embodiments, a phase change material is used that forms the crystal phase by crystal growth instead of using a nucleation mechanism. IST is one advantageous material in this regard.
As an aside, some Intel inventors submitted a patent application in 2013 (which remains unassigned) describing the challenges that are solved in the Ovonyx patent application cited above. The specimen in their particular example is a cross-point, 20 nanometer, dual-deck PCM. This exactly matches 3XP specifications!
Furthermore, there's also interfacial PCM ("iPCM" or "superlattice-like"), which further improves on all important attributes by reducing "entropic losses". Basically, the phase change material is layered thinly in order to reduce the amount of atomic movement that can take place.
Micron's most recent PCM patent application reveals this design:
Finally, it has been discovered that iPCM exhibits a magnetic mechanism, which really seals the deal versus all other next-gen memory technologies. While this magnetic mechanism is not exploitable at current fabrication sizes (think of this as trying to write with a pencil that is the size of a baseball bat - magnetic spin mechanisms are too subtle to leverage until we're under 10nm or so), there's no reason not to go all-in on PCM at this point. Toshiba outlines this well in a recent research paper whereby they propose using PCM (aka "PRAM") to replace DRAM and flash (the venerable hard drive survives in their model).
This money isn't being conjured - it is being siphoned from the DRAM market. Those manufacturers without a next-gen, nonvolatile DRAM replacement will die. The storage market (which will include NAND as well as 3XP) is going to be another $50 billion on top of this.
And that is all owned by Micron right now. While this is licensed to Intel through 2024, you can see the dire importance of a Micron acquisition, although I'm not sure that Intel is the high-bidder at this point.
As I outlined in the June article, standalone phase change memory does not represent the meat-and-potatoes implication of nanoscale Ovonic devices - logic does. The advent of nanoscale PCM means that Ovonic computing is now possible.
- Thus, we have shown that a three terminal OUM [Ovonic Universal Memory aka PCM] device can execute the logical operation "material implication" or IMP which combined with the logical false (or device reset), constitutes a computationally complete logic basis which means that all Boolean logic operations can be done using a single three terminal OUM device.
This is a semiconductor checkmate: a replacement for silicon-based computing that is better in all respects (speed, power, cost, density). This is just as big as the invention of the transistor itself (if not bigger, since this will take computing beyond binary and enable neuromorphic capabilities). As such, I believe that 3XP is a red herring that was put out there in order to distract us from this semiconductor mass-extinction event.
In the June article, I outlined that all of the intellectual property associated with Ovonic computing was stolen (here and here) in 2009, I have since learned that there were broad, "worldwide", royalty-bearing, cross-licensing agreements between ECD's memory and computing subsidiaries. Essentially, these agreements sub-licensed all current and future memory technology patents to the memory subsidiary - Ovonyx (which is now owned by Micron) while all computing patents were sub-licensed to the computing subsidiary - Ovonic Cognitive Computer, Inc ("OCC" hereinafter).
While OCC has no assets at this point, the contractual language is clear: the licenses are granted regardless of patent ownership. And the language is broad - perhaps too broad. Specifically, there is a binding clause on Ovonyx whereby "any rights and obligations shall apply to an affiliate", where "affiliate" is defined as "any entity which is owned or controlled by such party".
Since Micron now owns Ovonyx, they'd necessarily want to keep this a secret. And it was - until now. The sound that you are hearing is that of the entire semiconductor industry rushing to buy the OCC carcass from the ECD bankruptcy estate.
Most People Still Haven't a Clue
Computing capabilities aside, the $34 billion 2020 estimate from Intel for 3XP DIMMs partially reveals the Earth-shaking nature of this technology. What most people aren't yet realizing is that 3XP is "fast enough" to replace standalone DRAM in the vast majority of use cases. Research from 2011 outlines a hybrid PCM/eDRAM chip that increases performance and dramatically reduces power versus traditional homogeneous DRAM. This was done using assumptions from the old filamentary PCM technology: nano-PCM will improve the numbers even more dramatically.
From the patent applications, we know that the announced 128Gbit 3XP part is heavily sandbagged (i.e. - they don't want it to appear too disruptive). With the expected four planes (instead of two) and four bits per cell, that works out to exactly one terabit. This blows 3D NAND out of the water. This is the part that they were really going to announce - after the ECD bankruptcy had reached closure last summer.
But my article came along in the form of a giant monkey wrench, the ECD bankruptcy closing was further delayed and the Ovonyx CEO was deposed. Microsoft scrambled to put together a stop-gap Windows phone after it became apparent that they weren't going to be able to ship their PCM/3XP phone in the near future (this phone wasn't cancelled - just delayed until PCM is fully-secured).
Where the Sidewalk Ends
There's no incremental technology planned to succeed DDR4 DRAM. The industry is going to fragment at this next step (Micron's HMC, Samsung's WIO and AMD/nVidia/Hynix HBM). This transition is going to produce few winners and many losers. Who's going to win? The US Government has already weighed in on the matter (where "HMC" is Micron's "Hybrid Memory Cube"):
For those who haven't been following the supercomputing industry (perhaps all of you?), the US government has set a target performance of one exaFLOPS by 2020 with a power consumption limit of 20 megawatts. The challenges associated with this lofty goal are described in this 2013 report. An important note is that moving beyond DRAM is a foregone conclusion.
Everyone had assumed that the first supercomputer to reach this target would do so by leveraging only the minimum required "50 Petabytes" of memory. From page 14 in the report:
However, as shown in the Sandia slide above, Micron indicated that they can do 250 Petabytes of memory within the 20MW power constraint and that their hybrid memory cube is "expected to integrate well with local nonvolatile memory (whatever that turns out to be)" - sandbagging indeed. This blew everyone away when they revealed it in July. Have you ever seen a government presentation with a smiley face in it?
We're still waiting for the ECD bankruptcy to close (currently scheduled for January 20th) but I think that we're moving closer to an amicable settlement. It doesn't matter how much Micron and Intel have to pay for it - "believe that no other memory technology comes close".
Expect Micron to be acquired the instant after phase change memory has been legally secured.
Disclosure: I am/we are long MU, INTC, ENERQ.
I wrote this article myself, and it expresses my own opinions. I am not receiving compensation for it (other than from Seeking Alpha). I have no business relationship with any company whose stock is mentioned in this article.