When we think about our smartphones and other computing devices, memory isn't often the first feature we look at on the datasheet. Often the processor takes the spotlight, but memory is the real force behind the device's ability to get the job done. Flash has dominated the memory market as of late, but, as Moore's Law treads on, it faces some scaling issues, causing the industry to look elsewhere for memory solutions. Hewlitt Packard (HP)'s publicity of "The Machine" architecture has brought the term "memristor" back into the memory spotlight. This technology, also known as Resistive Random Access Memory (RRAM), is being researched and developed to become the next evolution in memory.

Non-volatile memory like flash is very important to all types of systems due to its ability to retain memory while turned off when not in use, thus saving energy – especially important in power-constrained embedded systems. But as applications push for faster and higher performance and lower power draw in smaller packages, memory companies are scouting RRAM's abilities to beat flash's performance as flash is nearing its scaling limits. Also, universities like Arizona State University are analyzing RRAM's strengths and weaknesses in various applications.

Comparing RRAM and flash

The concepts behind RRAM technology aren't new – they've been around since the 1960s[1], but have only gained significant interest in the last 10 years as a successor to current memory technology. Resistors, capacitors, and inductors are the three fundamental building blocks of electrical circuits, but memristors were a theorized fourth. A memristor is a resistor that can remember its history, thus functioning as memory, and RRAM is the technology that realizes this concept. A RRAM device can hold either a low or high resistive state depending on positive or negative voltages, respectively, which can be read as bits. And these states persist when disconnected from power, thus its potential as the next non-volatile memory technology.

ASU researchers have been active in developing RRAM technology. Professor Michael Kozicki is a pioneer in developing a type of RRAM – the Programmable Metallization Cell (PMC) and its commercial variant Conductive-Bridging RAM (CBRAM). Professor Kozicki and Associate Professor Hugh Barnaby have also been working on ways to make RRAM technologies useable in extreme environments like space, where the combination of low power and non-voltility is essential. Professor Sarma Vrudhula is an active supporter of RRAM technology use in new types of computation. Assistant Professor Shimeng Yu has been conducting research on RRAM since 2008. Yu says RRAM technology is faster (<10 ns) and programming voltage is smaller (<3 V) compared to current flash memory (>10 µs and >10 V).

RRAM is also poised to be more reliable than flash, Yu says. Memory reliability is judged on endurance (number of write cycles before loss of integrity) and retention (the readable lifetime of data). Non-volatile flash is lower in endurance compared to RRAM and can achieve 10^4 to 10^5 cycles. RRAM can achieve 10^6 up to 10^12 cycles. The typical retention standard for non-volatile memory is 10 years at 85 °C, which can be met by flash, and has the potential to be met by RRAM as well, Yu says.

The roadblock to RRAM's ability to be flash's successor in the near term is cost per bit. Flash is a very inexpensive technology to manufacture. Breakthroughs in 3D flash technology have further reduced the cost per bit of flash, delaying RRAM roadmaps for companies like SanDisk a few years until a cheaper, higher-yield manufacturing strategy can be developed for RRAM devices, Yu says. And the performance increases are not strong enough to overcome the cost increase to switching to RRAM.

A more brain-like memory

However, the memory market is not the only application for RRAM. Researchers are looking into "neurosynapse" applications, or making computers more like the brain.

Computing architectures today work in sequential operations. The CPU gets data from memory and makes a computation. But this often causes bottlenecks. The proliferation of data in today's applications is making people think of ways to process data in parallel like the brain does, Barnaby says. In the neural network of our brains, synapses connect active neurons in our brains as we learn. The idea is to use RRAM memory to function as a synapse between artificial neurons in the circuitry. This will be beneficial for applications like image recognition and speech recognition that involve some intelligence, Barnaby says.

With these exciting developments, it's an exciting time to be working with memory technology, which may soon steal some of the processors' spotlight.

[1] Hickmott, T. W. Low-frequency negative resistance in thin anodic oxide films. J. Appl. Phys. 33, 2669–2682 (1962). http://scitation.aip.org/content/aip/journal/jap/33/9/10.1063/1.1702530