Intel is Allegedly Shipping 14nm Products With MRAM

AlphaAtlas

[H]ard|Gawd
Staff member
Joined
Mar 3, 2018
Messages
1,713
Alternative non-volatile memory technologies like MRAM are a hot area of research, and some companies like Everspin are already producing MRAM products for niche use cases. However, a recent report by EE Times claims that Intel is shipping products made on a "22FFL" process with "the first FinFET-based MRAM technology." Intel itself didn't mention anything about their customers, and only describe the process as "production ready." Meanwhile, Samsung and Global Foundries say they introduced MRAM into manufacturing processes of their own. MRAM probably won't show up in consumer desktop products anytime soon, but the fact that it's (allegedly) shipping in some commercial 14nm product, and that competitors are taking an interest in it, is huge step forward.

In addition to being seen as a promising candidate for standalone devices to replace memory chip stalwarts DRAM and NAND flash - which are facing serious scaling challenges as the industry moves to smaller nodes - MRAM, which is a non-volatile memory, is appealing as an embedded technology replacement for flash and embedded SRAM because of its fast read/write times, high endurance, and strong retention... In its paper, Intel said that its embedded MRAM technology achieves 10-year retention at 200 Celsius and endurance of more than 10^6 switching cycles. The technology uses a 216 × 225 mm 1T-1R memory cell. Samsung, meanwhile, described its 8-Mb MRAM with endurance of 10^6 cycles and retention of 10 years.

Update 12/11/2018: Wikichip reports that Intel's 22FFL process is actually a "relaxed" version of their regular 14nm processes. The article was updated accordingly.
 
Last edited:
Was initially confused at how 106 cycles was an endurance leader over anything, and then after RTFA, it's 10^6. How about some superscript support, please, [H]?

From what they said about SRAM's current limitations, I look forward to what this will provide us in future CPU tech!
 
Was initially confused at how 106 cycles was an endurance leader over anything, and then after RTFA, it's 10^6. How about some superscript support, please, [H]?

From what they said about SRAM's current limitations, I look forward to what this will provide us in future CPU tech!
I actually had the same thought, but I didn't read the article. Thanks for this and correction.
 
Was initially confused at how 106 cycles was an endurance leader over anything, and then after RTFA, it's 10^6. How about some superscript support, please, [H]?

From what they said about SRAM's current limitations, I look forward to what this will provide us in future CPU tech!

Fixed, thanks. That's still a lot less endurance than SRAM or DRAM.

Did Intel ever post cell endurance claims for XPoint? I know they post endurance ratings for drives, but that doesn't really translate to individual cells, AFAIK.
 
Fixed, thanks. That's still a lot less endurance than SRAM or DRAM.

Did Intel ever post cell endurance claims for XPoint? I know they post endurance ratings for drives, but that doesn't really translate to individual cells, AFAIK.

Indeed, at 1TB/sec (L1ish) MRAM wouldn't last a fraction of a second before frying. It needs something like 10^15 cycles of endurance (same as DRAM), if it could do that and ship in storage-level capacities...oh my.

XPoint is supposedly 10^7 which is 100x higher than first generation enterprise SLC, I think.
 
Update 12/11/2018: Wikichip reports that Intel's 22FFL process is actually a "relaxed" version of their regular 14nm processes.

Relaxed. Not so tight around the bus. Give your nodes room to move.
 
Last edited:
Everspin MRAM is asininely expensive. If Intel can drive prices down that would be great. In my applications we'd replace all our EEPROM with MRAM if the prices could even be in the same ballpark.
 
But why? For the desktop this is expensive nonsense. Now, mainframe / supercomputer / datacenter usage could serve a purpose.
 
Everspin MRAM is asininely expensive. If Intel can drive prices down that would be great. In my applications we'd replace all our EEPROM with MRAM if the prices could even be in the same ballpark.

At work we replaced a bunch of EEPROM and battery backed SRAMs with Everspin parts. They are really nice and the applications aren’t cost sensitive. We are mostly afraid they will kill the parts or go out of business. They do deliver on their promises though.
 
The big Everspin was $35 back when I spun my ALU lookup tables to MRAM.
Price hasn't gone down much. Though better DIP adaptors are now available.
Wouldn't recommend the style I bought.

MR2A16ACYS35.jpg
 
Back
Top