The propeller heads in the world of high-performance computing ("HPC") live in an academic dream world and often forget that they've signed non-disclosure agreements. They've been discussing this openly for many years now - just use a small amount of DRAM in order to mitigate the downsides of PCM.
This appears to be Intel's upcoming Knights Landing HPC processor. Also relevant from this presentation is a slide that nicely illustrates that DRAM, like Flash, is at end-of-life beyond 16Gb densities:
...
3XP is a traditional lithography-intense memory technology: each layer needs to be fabricated individually so it is a lot more expensive than the new 3D NAND. In a four-layer stack of cells, you'd need to spend nearly four times the amount of time to create 3XP versus this BiCS implementation. In 2012, Samsung (OTC:SSNLF) outlined that 3D XPoint (yes, it used this term in 2012) wasn't ideal from a cost perspective:
If I'm spending $5 billion on a fab and depreciating half of that over three years, this works out to more than $26/second in overhead. Certainly, there's a lot of things happening in parallel, but you get the picture: equipment depreciation is the dominant cost of a semiconductor.
...
There's no reason to produce 3D NAND or 3D XPoint if you have the rights to produce BiCS PCM. I don't believe that Intel has those rights and, as such, has only a short-term solution with 3D XPoint. It will prove very costly, but I believe that Micron and Toshiba will prevail in the long term with their BiCS PCM, which can be produced in much greater density (think 64 layers instead of two or four). Most recently, Toshiba appears to be prepping for this change. In the distance, Samsung is building a $23 billion fab just as DRAM and NAND are hitting their end of life. I have nothing here, but I find this curious and see a potential partnership.
Don't get me wrong: Intel is going to do just fine with its walled garden packaged processor/memory/storage products. And it has many aspects of transactional memory locked up. This alone cost it a fortune - it's years ahead. And the company is doing a great job of playing a high-stakes game of chicken against Micron (see the MU share price, for reference).
...
I will say that I see lots of ECD technology that is awaiting the closure of its bankruptcy. Phase change window film, e-paper and display technology, for example, were recently announced by Oxford and Exeter spin-off Bodle Technology.
ECD's chalcogenide display technology - now owned by Micron - will obsolesce current LCD, LED, OLED and ePaper displays - a market of $80 billion annually. A current 4K television has just 8.3 megapixels due to the tremendous limitations of thin-film transistors. But ECD's Ovonic Quantum Control device solves all of those limitations, albeit creating a definite bandwidth issue, building displays with billions of pixels will become a possibility. Or smartphone displays that transform from backlit to color ePaper when you're out in the sunlight.
...
And then there is Ovonic optical routing. With today's technology, it takes about 75 milliseconds to get data from Los Angeles to New York. The vast majority of the 3,000-mile trip is performed over fiber optic cable. The speed of light is 186,000 miles per second so, in an all-optical situation, this trip should require less than 2 milliseconds.
There's more than 70 milliseconds of overhead in optical-electrical-optical conversion. That is, in order to traverse thousands of miles, the electrical data needs to be converted to optical. But, at each "hop" (i.e. - every major city), it needs to be converted back into electrical so that it can be routed by an electricity-based processor. With Ovonic optical routing, this does not need to happen. The world is about to get a lot smaller from a communication perspective. And Micron now owns the fundamental patents on this technology.
I think it is worth reading the article (and look at some of the figures). There is a lot of speculation involved, so take it with a grain of salt, but it still contains some very interesting information.
