Researchers suggest using stochastic MRAM elements to create highly efficient AI neural network devices

Researchers from Northwestern University developed a new method of building artificial neural networks using MRAM-based stochastic computing units. The researchers say that this design could enable AI devices that are highly energy efficient.

MTJ-based stochastic computing unit structure (Northwestern University)

Embedded MRAM technologies are being adopted at major foundries, which enable the use of these technologies for unconventional computing architectures that use the stochasticity of MRAM cells (rather than their nonvolatility), to perform energy-efficient computing operations. MRAM cells exhibit stochastic switching characteristics, which is a challenge for reliable memory devices. But for neural networks, this can be taken advantage of if the MTJs are appropriately designed.

Lucid Motors adopted Everspin's MRAM in its Lucid Air all-electric sedan

Everspin Technologies announced that Lucid Motors has designed-in the company's 256 Kb MRAM into its master powertrain system for its Lucid Air all-electric luxury sedan.

Everspin Technologies chip photo

Everspin's The MR25H256AMDF MRAM device is designed for automotive applications and is qualified to the AEC-Q100 Grade 1 standard for use in demanding memory applications that require extreme reliability in critical data capturing systems.

Avalanche announces space-grade Gigabit-density STT-MRAM

pMTJ STT-MRAM developer Avalanche Technology announced its third-generation 1Gb space-grade parallel asynchronous x32-interface high-reliability P-SRAM (Persistent SRAM) memory devices. The company says that these new devices enable customers to design unified memory architecture systems for high reliability aerospace applications, in extremely small form factors.

Avalanche pMTJ STT-MRAM P-SRAM Serial QSPI Evaluation Kit photo

Avalanche's 2nd-Gen P-SRAM evaluation kit

The new Parallel x32 Space Grade series is offered in 512Mb, 1Gb, 2Gb and 4Gb density options and has asynchronous SRAM compatible 45ns/45ns read/write timings. Data is always non-volatile with >10^14 write cycles endurance and 10-year retention at 125°C. All four density options are available in a small footprint 142-Ball FBGA (17mm x 11mm) package.

Avalanche starts production of space-grade 16-64Mb STT-MRAM devices

pMTJ STT-MRAM developer Avalanche Technology announced that it is now shipping new space-grade parallel asynchronous x16-interface high-reliability P-SRAM (Persistent SRAM) memory devices, based on its latest STT-MRAM technology.

Avalanche pMTJ STT-MRAM P-SRAM Serial QSPI Evaluation Kit photo

Avalanche says that its STT-MRAM devices are smaller and more efficienct compared to Toggle MRAM based products, currently adopted in aerospace applications. The Parallel x16 Space Grade series is offered in 16Mb, 32Mb and 64Mb density options and has asynchronous SRAM compatible 45ns/45ns read/write timings. All three density options currently in production and available within industry standard lead times.

Gyrfalcon's new AI chip first to use TSMC's embedded MRAM

In June 2017 it was reported that Taiwan Semiconductor Manufacturing Company (TSMC) will start producing embedded MRAM in 2018 using a 22 nm process. In what may bet he first adoption of TSMC's eMRAM technology, AI accelerator startup Gyrfalcon Technology announced the commercial availability of its LightSpeeur 2802M, AI ASIC that include TSMC's eMRAM.

The 2802M ASIC has 40MB of eMRAM memory, which can support large AI models or multiple AI models within a single chip. Applications include image classification, voice identification, voice commands, facial recognition, pattern recognition and more.

SMART Modular Technology launces aMRAM-enhanced n nvNITRO U.2 Sotrage Accelerator

SMART Modular Technologies has launched its new nvNITRO U.2 Storage Accelerator that features Everspin's STT-MRAM technology. The nvNITRO is ideally suited for synchronous logging applications such as those used for financial trading.

Smart nvNITRO U.2 MRAM card photo

SMART's nvNITRO U.2 Storage Accelerator uses a standard NVMe interface that is 1.2.1 compliant and provides less than six microseconds of industry-leading low latency access with persistence so that all logging data is safe. The U.2 form factor brings with it the advantage of being hot-swappable.

IBM introduces its latest 19.2TB enterprise SSD drivers with Everspin's STT-MRAM

A couple of days ago, Everdisplay disclosed its first major design win with a "top enterprise storage vendor" for its 40nm 256Mb STT-MRAM chips. We now know that this vendor is IBM - as it introduced its latest-generation enterprise SSD FlashSystem, which indeed includes Everspin's STT-MRAM.

IBM 2018 FlashSystem spec slide

Using MRAM instead of DRAM memory enabled IBM to remove the relatively large supercapacitors (used to make the DRAM non-volatile) and so the company was able to reduced the size of its drives and switch to a standard 2.5-inch U.2 drive form factor. The new FlashSystem SSDs support up to 19.2TB of 64L 3D TLC NAND. IBM's system uses a 20-channel NAND interface and a four-lane PCIe 4.0 host interface that can operate in dual-port 2+2 mode.