IBM and development partner Samsung announced they've developed a process to manufacture a type of non-volatile RAM that is up to 100,000 times faster than NAND flash and never wears out.
The two companies collaborated to develop next-generation magnetoresistive RAM (MRAM) using spin-transfer torque (STT) technology, which would lead to low-capacity memory chips for Internet of Things sensors, wearables and mobile devices that currently use NAND flash to store data.
IBM and Samsung published a paper in IEEE Magnetics Letters outlining how they scaled STT MRAM down to 11 nanometers for the first time by using 10 nanosecond pulses and just 7.5 microamperes — "a significant achievement."
NAND flash on average takes one millisecond to write data compared to MRAM's 10 nanoseconds -- meaning MRAM is 100,000 times faster than NAND flash on writes and 10 times faster on reads, said Daniel Worledge, the senior manager of MRAM development at IBM Research, in an email reply to Computerworld.
"This is important because it now falls into the sweet spot compared to other memory technologies and this level makes it viable to manufacture," Worledge said.
"This could never be done with in-plane magnetized devices — they just don't scale," Worledge said, referring to an earlier version of Spin Torque MRAM. "While more research needs to be done, this should give the industry the confidence it needs to move forward. The time for Spin Torque MRAM is now."
Spin-torque MRAM can be used for a new type of working memory in ultra-low power applications. For example, it can be used in IoT or mobile devices, where it uses very low power when it's on and storing information, and when it's not actively being used, it uses zero power because it's not volatile.
Worledge said he doesn't believe that IBM's STT MRAM will replace DRAM anytime soon, but he said it can easily replace embedded flash, since MRAM is easier to embed, is faster and has unlimited reads and writes.
"If you look at flash memory, which is what you use in your digital camera, it can only be written about 10,000 times before it wears out. That's fine for taking photos, but if you were to use it as a working memory, it would wear out in less than a second," Worledge said.
MRAM doesn't wear out because spin torque technology uses a tiny current to switch a bit from a zero to a one and vice versa. Data is stored as a magnetic state versus an electronic charge, providing a non-volatile memory bit that doesn't suffer wear-out or data-retention issues associated with NAND flash technology.
Unlike NAND flash, spin-torque MRAM technology transistors don't need to be erased first before being rewritten with new data, which also greatly simplifies chip design and reduces overhead.
IBM, with a partner, plans to optimize the MRAM's engineering parameters for mass production in as little as three years. IBM has expertise in magnetic materials and devices, and in MRAM circuit design, which will accelerate the path to products for its partners.
IBM has been working on developing MRAM chips for 20 years. Prior to the latest development, the company could not scale down the technology enough to make the chips financially viable for manufacturing.
In a new blog, IBM stated that "MRAM is an ideal technology for always-on devices, such as IoT sensors, mobile devices and wearable electronics, because it offers more storage and longer battery life. In addition, since MRAM uses standard transistors, and is compact and robust, it's more easily embedded on the same chip as logic and other functions, compared to flash memory. Therefore many semiconductor foundries are considering replacing embedded flash with embedded STT MRAM at the 28nm size and below."
There are new memories technologies in various stages of commercial development that are poised to replace NAND flash as a non-volatile storage medium - meaning data remains even after the power is turned off.
Among the top candidates being developed are Memristor, or Resistive Random Access Memory (ReRAM), Phase Change Memory (PCM), Ferroelectric Random Access Memory (FeRAM), and MRAM. Out of the list of new memories, only FeRAM and MRAM account for "a reasonable market share and they are quite commercialized in the market," according to a report by MarketsandMarkets.
IBM is not alone in its development of MRAM.
In 2011, Hynix Semiconductor and Toshiba Corp. formed an MRAM development partnership. That same year, Samsung acquired Grandis Inc., a Milpitas, Calif., maker of STT MRAM technology. At the time, however, the technology was too expensive to mass produce. That changed as chips were miniaturized and processes become more standardized.
In April, semiconductor maker Everspin Technologies announced it was shipping samples of the industry's most dense MRAM chip, which could replace standard DRAM for write-caching operations. The chip, with up to 1 gigabit of capacity, was the company's third generation of MRAM and is aimed at replacing persistent DRAM on servers and storage arrays.
In essence, Everspin's MRAM could act as the first tier of storage in a storage array or server, protecting data not yet stored on a mass storage device, such as NAND flash or hard disk. Everspin's MRAM uses a 40nm and 28nm size transistor compared with IBM's new 11nm cell.
One of the largest markets for higher speed, lower power and longer lasting non-volatile memories like MRAM is the IoT The IoT includes vehicles, buildings and the sensors within them, which are enabled by network communications governed by computers. For example, smart thermostat systems use Wi-Fi for remote monitoring and can detect when people are in certain rooms and adjust temperatures accordingly to save power.
The IoT market alone is expected to grow from $1.9 trillion in 2013 to $7.1 trillion in 2020, according to industry research firm IDC.
Because IoT systems are often powered with batteries or are constantly monitoring and communicating, it's important that the power they consume remains low.
IBM and Samsung's MRAM sips power; it's able to write a bit of data with just 7.5 microamps, which allows IBM to use very small transistors enabling a very dense chip. Its write-error rates are also exceptionally low. For every billion bits (gigabit) written, there is less than one error, Worledge said.
Previously, the only MRAM that could achieve such as low error rate were produced using 100nm process technology, Worledge said, ten times larger than IBM's transistors.
"MRAM can be scaled to very small dimensions. The basic building block - a magnetic tunnel junction - can be made as small as 11 nanometers in size," Worledge said. "That means we could make very dense and fast MRAM chips in the future that can be used as fast cache memory in IBM servers."
Eleven nanometers is about 10,000 times smaller than the diameter of a human hair.
Worledge said he could not comment about when an MRAM product would reach the market. "IBM's role is to do research and development to lead the way to new products, which our partners will manufacture," he said.