Resistive random-access memory

So when you turn it off and back on, it resumes from where it was stopped, not to restart from the beginning.

And that's where it comes the effectiveness of RRAM or MRAM into play.

We can turn things off very quickly, turn it back on in microseconds, granularity, and preserving the system state that's actually orders of magnitude faster if you want to do the same thing with flash memory devices.

So this is where the non-volatile memory, MRAM and RRAM, becomes really handy.

So to keeping the state of your processing system where it is, you turn it off, wait for the data to be fed from the sensors, and then you can turn the old system again.

You can run your AI workload.

What are your thoughts on this, Craig? Do you think this is the future or potential future of memory?

Well, the semiconductor industry has pursued the universal memory for decades through MRAM, re-RAM, phase change memory, and other candidates.

A challenge rarely demonstrating an interesting cell is achieving competitive density, reliability, yield, cost, and interface compatibility simultaneously, which is what we need to see.

This is very interesting.

Right.

This is a prototype of an ReRAM chip.

Which stands for resistive random access memory.

And for those unfamiliar, traditional memory stores data by trapping electrical charges.

Right.

ReRAM works differently.

It stores information by physically changing the electrical resistance of a solid dielectric material.

It's a completely different approach.

It reaches three hundred forty-six point six tops per watt for one bit by one bit multiply accumulate operations with eighty-five point two percent inference accuracy at sixteen row parallelism and four-bit ADC resolution, and eighty-nine point one percent at eight row parallelism with three-bit resolution.

And FeFETs have a structural advantage over RRAM here, a higher on/off ratio and lower read and write energy because the read is a gated transistor current rather than a resistive divider that has to push real current through the storage element itself.

But look closely at those accuracy numbers, eighty-five to eighty-nine percent because they're the tell.

Analog compute in memory isn't limited by the array.

GlobalFoundries and Fraunhofer IPMS have that aluminum co-doped HZO Ferram integrated into twenty-two FDX sub one volt, eight nanometer film, nanosecond switching, retention at a hundred and fifty degrees.

Integrated in a production flow and prize-winning in June twenty twenty-six.

But remember, GF's actually shipping twenty-two FDX plus embedded non-volatile memory is RRAM, not ferroelectric.

Samsung has the FeFET NAND Nature paper, five bits per cell, near zero pass voltage, ninety-six percent string power reduction.