Memory Is Holding Up the Moore's Law Progression of Processing Power

Processors are getting faster and faster, but a processor is worthless if it's not being fed information fast enough.

Image: DRAM/Wiki

Computer memory has reached something of a crisis state. You'll mostly hear otherwise, however, as hard-drive memory becomes cheap enough that it might as well be free, while the big names behind the current cloud storage push gleefully lower that price to actually free.

But this sort of memory, secondary memory, has really fuck all to do with how your computer actually computes. The processing done by a computer is dependent not on how many torrented Blu-ray files you can shove onto a solid-state platter, but on how fast a processor can access memory. This is the crisis: actual microprocessors are getting faster, just as we'd expect from Moore's Law, but that means nothing for the actual speed of future computers.

Simply, memory is becoming a critical bottleneck in the Moore's Law progression of processing power. Our "old" memory methods are stagnating, relatively, compared to the rapid advance of bare processing power. The limit is the simple physics of electrical charge, and how we're able to store that charge in such a way to maintain and retrieve bits of information. If rapid advances aren't made soon to correct this, the result will be future processors humming awaybut without anything to actually process. 

This is one of the primary problems that spintronics hopes to solve, the advancement of random-access memory. SRAM and DRAM are the primary components of this memory in most machines now, with the prior being faster, more expensive, and less dense in terms of information storage. SRAM holds only the most immediate of the immediate data a processor needs, with the latter making up the rest of the immediate memory. SRAM is the stuff sitting just right in front of a microprocessor, while DRAM makes up the stuff that should be right in front of a microprocessor but isn't, owing to the cost/space limitations of SRAM. (Even before SRAM in line, there are miniscule registers that hold only relative handfuls of bits at a time.) 

A computer's hard drive and other secondary memory devices like USB sticks might as well exist on another planet from actual computing. Microprocessors have no direct access to this kind of memory, which is delivered to RAM memory by different programs, where it becomes of the active computing landscape. The dream is of a single memory system, with the speed of SRAM but the storage density of nothing yet created.

Recent years have seen the development of a third kind of RAM, MRAM. MRAM uses spintronics to store information. Spintronics is a technology that ditches the property of electrical charge, the red blood cells of technological civilization, in favor of a different particle property called spin. Spin is about what it sounds: a measure of a particle's angular momentum as it rotates around, creating a very tiny magnetic field. Particles have different spin "states"; electrons can be found with either an "up" or "down" orientation (or, as allowed by quantum mechanics, a superposition between the two, but that's a concept for later).  

Specifically, an MRAM device orders tightly packed rows of electrons within a "sandwich" of two materials, with one of those materials containing electrons with variable spin orientations and the other having a constant polarization. The bit of information is determined by whether neighboring electrons, one in the variable plate and one in the constant plate, have the same spin or different spins. 

MRAM has had a fundamental barrier of its own, however. Current MRAM uses something called tunneling magnetoresistance, a mechanism taking advantage of the quantum mechanical phenomenon of tunneling, in which electrons can "tunnel" through some insulating material thanks to quantum mechanical effects. The relative polarizations of the two plates leads to different probabilities that electron tunneling will occur between them, and, thus, it's possible to determine the relative polarizations by "watching" the fields between the plates. So: information.

Current MRAM technologies impose a size limit for the bit-storing "cells" of around 100 nm. Past that, the cells begin to overwrite their neighbors and, what's more, the current schemes take a lot of juice to write data. It's not an ideal technology, and currently its applications are limited to industrial uses, where holding bits of information in RAM memory without the need for a constant power supply (one general advantage or MRAM) is highly desireable.   

A once-thought-abandoned line of alternative MRAM research involves something called the giant magnetoresistive effect (GMR). A RAM memory read/write scheme using this principle would operate using current "injections" of polarized electrons at angles perpendicular to the memory plate surfaces. The information bit is "read" by a memory head that delivers this current: If the electrons have the same spin, resistance to the current is minimal; if they have different spins, resistance is greater. So, electric charge scooting around a computer at two-thirds the speed of light are still part of the MRAM scheme, but it's not needed for holding the actual inormation. The ideal is that they could hold all of the computer's information though, in one location in view of the processor. This is where the speed advantage comes from; there's no longer any need for time-consuming address lookups in DRAM. It'd be like the contents of a storage unit fitting into a backpack.

The limit on this alternative technology has come from the fact that achieving the 100 percent polarization preferred for the reading/writing current is difficult at room temperature, so large read/write heads have been impossible. The technology was even thought to be abandoned, at least until a couple of weeks ago, when a team of researchers based at Johannes Gutenberg University published a conspicuously underreported study showing 100 percent polarization in a memory read/write head made from half-metal materials known as Heusler alloys.

"This class of materials has long been under investigation and there is substantial theoretical evidence for the required electronic properties of Heusler compounds but no single experiment has previously been able to confirm 100 percent spin polarization at room temperature," said the study's lead author, Martin Jourdan, in a JGU press release.

"[Replacing tunneling-based heads with GMR heads] will reduce the device resistivity, which is very important for increased storage density, i.e. smaller read head areas," explained Jourdan in an interview. "However, the GMR is smaller than the TMR [tunneling-based head], which is why GMR read heads were replaced by TMR read heads. But with a huge spin polarization, by using Heusler materials, the GMR will become potentially much larger." And so MRAM finds a new-old possibility for beating its own prohibitive limits.

Jourdan noted that he's just a physicist and not an engineer, so anticipating a bit-storage cell size would be out of his range. It's certainly a step toward memory technology's true promised land, which is not actually the cloud, but in single component storage systems, ones local to your personal machine so long as that machine contains its own processor.

Such super-advanced MRAM systems would effectively trash what's known as the memory hierarchy, the pyramid of storage technologies that extends from infinitesimal storage registers within actual processors to super-fast but expensive and big SRAM to less expensive but slower DRAM to actual hard-drives. The speed limits involved in scooting data among these distinct systems is what is poised to break Moore's Law for the first time.

This is what the gap looks like and you'll probably find a version in most computing systems textbooks: 

Source: Synopsys 

It's a fascinating problem: processors get faster and faster, but a processor is nonetheless worthless if it's not being fed information fast enough. Future MRAM units with the ability to keep every memory system together within "view" of a processor will potentially flip the table and, really, there aren't too many other solutions beyond spintronics, besides slower computers, that is.