Intel's FinFET-Based Embedded MRAM is Ready for Production

[Raevenlord]

A report via EETimes slates Intel's own working MRAM (Magnetoresistive Random-Access Memory) is ready for production in high-volume manufacturing. MRAM is a nonvolatile memory technology, meaning that it retains information even if there is a change in powerstate (ie, power loss), meaning that it's more akin to a storage device than to, say, RAM.

But why does MRAM matter, really? Well, MRAM is being developed as a long-term candidate to a universal memory solution, replacing both DRAM (a volatile memory technology) and NAND flash (a nonvolatile one), since node scaling with these technologies is becoming increasingly harder. MRAM promises better-scaling (at the foundry level) processes, with much higher yield rates. The fact that MRAM has been demonstrated to be able to achieve 1 ns settling times, better than the currently accepted theoretical limits for DRAM, and much higher write speeds (as much as thousands of times faster) compared to NAND flash.





MRAM, at its current production characteristics, enables 10-year data retention at 200°C, and endurance of more than 10^6 switching cycles, said Ligiong Wei, an Intel engineer who presented the paper Tuesday. In addition to high endurance, it's being reported that the 22-nm embedded MRAM technology has a bit yield rate that's greater than 99.9% - an amazing feat for such a relatively secretive technology.

There seem to be some questions regarding what exactly is the production process being used a 22 nm process is being reported, but other outlets say this is actually a "relaxed" 14 nm fabrication node with much of Intel's 14 nm technologies thrown in. Whatever the case, it's apparently ready for prime-time.

View at TechPowerUp Main Site

 

[Rictorhell]

A very exciting new technology that seems to hold the potential for new types of computers and devices and probably has a lot of unexpected and convenient uses that no one has thought of, yet. I can't wait to see the first devices to use this technology and how they compare to current hardware.

 

[TheLostSwede]

If I'm reading the specs and the die photo correctly, that die has about 7MB worth of MRAM in total. What's crucially missing, is how many of those die that goes into one chip. Without knowing this, is hard to tell if this is something that could go into mainstream hardware as soon as there's support for MRAM, or of it's a niche product that might never scale in terms of memory size.

 

[silentbogo]

and endurance of more than 106 switching cycles

10^6

 

[Raevenlord]

10^6

Merci =)

 

[notb]

If I'm reading the specs and the die photo correctly, that die has about 7MB worth of MRAM in total. What's crucially missing, is how many of those die that goes into one chip. Without knowing this, is hard to tell if this is something that could go into mainstream hardware as soon as there's support for MRAM, or of it's a niche product that might never scale in terms of memory size.

MRAM is fairly similar to DRAM physically and will have a similar density. Don't expect a boost in memory volume (but you have 3D XPoint for that).
The main advantage of MRAM is power saving. It uses just few % energy DRAM needs.
Theoretically it should also be significantly faster.

 

[mtcn77]

Intel is about to score a pretty penny in the server space if the energy savings are palpable.

 

[notb]

Intel is about to score a pretty penny in the server space if the energy savings are palpable.

Well, I'm not so sure about servers in general. It's just a question of energy cost and RAM isn't the most power-hungry element. It may also be a fairly expensive tech at first - quite possibly covering the energy savings.
But it can have a serious impact on mobile devices (especially tablets and smartphones). A moderate amount of RAM can suck as much as the SoC under load (~2W).
It's not a coincidence that Samsung is so active in MRAM development.

 

[mtcn77]

Well, I'm not so sure about servers in general. It's just a question of energy cost and RAM isn't the most power-hungry element. It may also be a fairly expensive tech at first - quite possibly covering the energy savings.
But it can have a serious impact on mobile devices (especially tablets and smartphones). A moderate amount of RAM can suck as much as the SoC under load (~2W).
It's not a coincidence that Samsung is so active in MRAM development.

Not in general rankings, but dram recycles and that forms the bulk of its charge.

 

[PanicLake]

Yay! Finally we'll have malware that won't go away after a restart!

 

[Zubasa]

So, finally RamBus @INTEL is actually doing work for once.
Instead of patent trolling everyone and their grandma.

 

[dogsbody]

Yay! Finally we'll have malware that won't go away after a restart!

There is no technical reason that would make reinitialization of such NVRAM on power-up impossible. In fact same applies to DRAM. Boot will always begin with less a volatile memory ("ROM", flash nowadays).

 

[notb]

So, finally RamBus @INTEL is actually doing work for once.

MRAM has been in development for 20 years (this particular approach). It's not just Intel, RamBus and Samsung. It's also a hot topic on universities etc.
But searching for a non-volatile alternative to RAM started almost immediately after DRAM became popular. The fact that DRAM needs to be refreshed all the time is really idiotic, but we didn't have a better technology back then.

Lights in corridors and stairways usually turn off after few minutes.
Imagine you have a light source at home which does the same, so you have to go to the switch or at least wave your hand every 5 minutes. Imagine finding a solution to this takes few decades...

 

[silentbogo]

MRAM has been in development for 20 years (this particular approach). It's not just Intel, RamBus and Samsung. It's also a hot topic on universities etc.

I think carbon nanotube based NRAM is probably a more viable solution at this point. Nantero already have working prototypes, and Fujitsu had recently acquired a license for mass production. They even went as far as developing a drop-in DDR replacement based on new tech, just to demonstrate the viability and scalability. I'm talking current-level package density.
Intel's MRAM is more of a high-speed alternative for embedded products, and I'm not aware if there were any plans or announcements made regarding consumer segment or replacing conventional RAM. Even Intel's limited amount of news and slides only points that it's gonna be a non-volatile alternative for microcontrollers & RF chips (at least for now).

 

[notb]

I think carbon nanotube based NRAM is probably a more viable solution at this point. Nantero already have working prototypes, and Fujitsu had recently acquired a license for mass production. They even went as far as developing a drop-in DDR replacement based on new tech, just to demonstrate the viability and scalability. I'm talking current-level package density.

NRAM is still an experimental product developed by a small company with a good idea. Fujitsu license for production doesn't mean they will start making it anytime soon.

NRAM uses CNTs (carbon nanotubes). At this point we don't even have a method for mass manufacturing CNTs that would be usable in electronics (not to mention making it cheap). You don't build them. You grow them.

We know graphene is awesome and it'll surely appear in many things we use. But don't hold your breath.

Intel's MRAM is more of a high-speed alternative for embedded products, and I'm not aware if there were any plans or announcements made regarding consumer segment or replacing conventional RAM. Even Intel's limited amount of news and slides only points that it's gonna be a non-volatile alternative for microcontrollers & RF chips (at least for now).

But unlike NRAM, MRAM is around the corner. We may see it in PCs and phones later this year.
MRAM is also based on technology large semiconductor companies know very well. It's a normal chip that can be made in existing fabs.

NRAM is likely a better idea. But it's cutting edge physics and material science. Developing new technologies takes years or decades.