Crossbar Pushes Resistive RAM Into Embedded AI

Deal with Microsemi and foundries means its nonvolatile embedded memory can be integrated into the most advanced chips

2 min read
Crossbar's demo unit.
Photo: Crossbar

Resistive RAM technology developer Crossbar says it has inked a deal with aerospace chipmaker Microsemi, allowing the latter to embed Crossbar’s nonvolatile memory on future chips. The move follows the selection of Crossbar’s technology by a leading foundry for advanced manufacturing nodes. Crossbar is counting on resistive RAM (ReRAM) to enable artificial intelligence systems whose neural networks are housed within the device rather than in the cloud.

ReRAM is a variant of the memristor, a nonvolatile memory device whose resistance can be set or reset by a pulse of voltage. The variant that Crossbar qualified for advanced manufacturing is called a filament device. It’s built within the layers above a chip’s silicon, where the IC’s interconnects go, and it’s made up of three layers: from top to bottom—silver, amorphous silicon, and tungsten. Voltage across the amorphous silicon causes a filament of silver atoms to cross the gap to the tungsten, making the memory cell conductive. Reversing the voltage pushes the silver back into place, cutting off conduction.

“The filament itself is only three to four nanometers wide,” says Sylvain Dubois, vice president of marketing and business development at Crossbar. “So the cell itself will be able to scale below 10 nanometers.” What’s more, the ratio between the current that flows when the device is on and when it is off is 1,000 or higher. Competitors for a spot as an embedded nonvolatile memory, such as magnetic random-access memory (MRAM), are orders of magnitude smaller.

With the foundry backing, Crossbar is hoping to market embedded ReRAM as a key to moving artificial intelligence systems into small or mobile devices such as surveillance cameras or drones. “AI is going to the edge,” says Dubois. “You cannot rely on the cloud for assisted driving or autonomous driving or even for a mobile phone.”

“The biggest challenge facing engineers for AI today is overcoming the memory speed and power bottleneck in the current architecture to get faster data access while lowering the energy cost,” Dubois said in a press release. Embedding ReRAM in a processor should give that fast access and energy savings, he argues. The company is demonstrating a test chip at the Embedded Vision Summit next week in Santa Clara, Calif. It’s capable of running face recognition and license plate recognition, says Dubois. And it can train to recognize new faces without help from the cloud.

The Conversation (0)

3 Ways 3D Chip Tech Is Upending Computing

AMD, Graphcore, and Intel show why the industry’s leading edge is going vertical

8 min read
A stack of 3 images.  One of a chip, another is a group of chips and a single grey chip.
Intel; Graphcore; AMD

A crop of high-performance processors is showing that the new direction for continuing Moore’s Law is all about up. Each generation of processor needs to perform better than the last, and, at its most basic, that means integrating more logic onto the silicon. But there are two problems: One is that our ability to shrink transistors and the logic and memory blocks they make up is slowing down. The other is that chips have reached their size limits. Photolithography tools can pattern only an area of about 850 square millimeters, which is about the size of a top-of-the-line Nvidia GPU.

For a few years now, developers of systems-on-chips have begun to break up their ever-larger designs into smaller chiplets and link them together inside the same package to effectively increase the silicon area, among other advantages. In CPUs, these links have mostly been so-called 2.5D, where the chiplets are set beside each other and connected using short, dense interconnects. Momentum for this type of integration will likely only grow now that most of the major manufacturers have agreed on a 2.5D chiplet-to-chiplet communications standard.

Keep Reading ↓ Show less