Flash memory has revolutionized data storage and systems technology in the last decade. But flash has serious performance, power, and endurance problems. And it doesn't shrink the way other semiconductor technologies do, so vendors have been forced to go to 3D fabrication with large feature sizes.
That's why the industry is excited about the NVRAM technologies coming into play. Intel's 3D Xpoint is the most famous, but several other companies are beavering away to create viable alternatives to static RAM (SRAM), and, for the win, DRAM.
What is MRAM?
As described in the surprisingly clear PSC patent MRAM is a
. . . non-Volatile memory technology that stores data through magnetic storage elements. These elements are two ferro-magnetic plates or electrodes that can hold a magnetic field and are separated by a non-magnetic material. . . . [O]ne of the plates has its magnetization pinned (i.e., a "reference layer"). . . . The second plate is typically referred to as the free layer and its magnetization direction can be changed by a smaller magnetic field. . . .
Here's a simple graphic:
Magnetic RAM, or MRAM, has been in production for a decade, used mostly in embedded systems where flash can't cut it, usually for power or endurance reasons. Vendors, such as Everspin and Spin Transfer Technologies, have been working to increase the density, performance and endurance of MRAM.
MRAM has some powerful advantages over flash, including:
- Low-cost manufacturing
- Scales well below 28nm, unlike flash
- 100s to 1,000s less power for writes than flash, and no refresh power like DRAM
- Byte addressable, like DRAM and unlike flash
But all is not glorious in the MRAM world. Many embedded systems must operate in high temperatures, which hurts data retention endurance. While MRAM cells take a fraction of the area of SRAM, today it isn't as fast, nor is the retention and endurance as good. And performance is also a problem when MRAM is competing with DRAM, as well as density.
What is Precessional Spin Current?
The physical implementation of PSC is three thin layers on top of a perpendicular magnetic tunnel junction (pMTJ). When I spoke to STT they didn't want to go into what the layers did, beyond that fact that the layers are cheap to add - $1/wafer - and that they are licensing the technology to other MRAM makers.
But what these thin, patented layers do to a pMTJ is fourfold:
- Reduces the current required to write to the cell
- Raises the cell's data retention ≈10,000x when being read, reducing read disturb problems
- Since the write current is reduced, the PSC increases endurance 100x to 1000x
- And, cell behavior improves as feature size shrinks
Hyperbolic claims are the bread and butter of most startups, so I drilled down into how STT know their technology delivers the goods. I came away impressed.
STT has a magnetics R&D fab with a five day cycle time so they can quickly build test chips. They subject the test chips - each with thousands of pMTJ devices - to extended testing for temperature, switching current, and other parameters.
Translation: this is real hardware, not a web-service Powerpoint. Yes, it's in a lab today, but STT knows what is needed to take products to market.
The Storage Bits take
STT has raised over $130m to fund their R&D. They have a top notch technical team.
MRAM is currently being shipped by TSMC, Samsung, and Global Foundries. And companies such as SK Hynix, Toshiba, and Qualcomm are also working on the technology, aided by $1B+ industry investment fab equipment, materials research, and EDA tools.
So what will MRAM uptake look like? I expect an accelerated version of what NAND flash went through. Flash started shipping in the 80s, and slowly grew volume in niches that disk technology could not serve, such as Compact Flash. And then, the tipping point: flash reached DRAM prices. Suddenly volumes spiked, prices dived, and over the next five years flash became an overnight success, after close to 20 years.
STT's PSC tech won't reach commercial products until mid-to-late 2019, and the mostly for embedded applications in autos, geo-physical, and other difficult environments. But the march is on, and within a decade I expect to see MRAM competitive with DRAM, especially considering power. In an increasingly mobile world, that will be an excellent outcome.
Courteous comments welcome, of course.