We’ve been hearing about commercially feasible memristors—a type of electrical component that’s persistent like flash memory, but as fast as DRAM to access—for quite some time now. Most prominently, HP got its proof of the memristor’s existence published in Nature back in 2008, and it partnered with SK Hynix to produce memristors in 2010, but there’s been little new fruit on that front since (aside from the HP Labs project called "The Machine," which has since been redesigned around more conventional technologies). Now, a new startup called Knowm is taking its own shot at bringing the memristor to market, this time as part of a machine-learning solution.
At the hardware level, Knowm is taking a different tack than HP’s "crossbar" memristor design, according to this EE Times article. Instead, the company builds memristors using a technology it calls "Thermodynamic RAM," which appears to be fabricated using a somewhat more traditional process. Thermodynamic RAM is claimed to be easily scalable, and Knowm plans to offer its tech as an add-on fabrication step in its own facilities for CMOS ASICs sourced from other chip makers.
Knowm is using memristors to implement a form of computing architecture it’s calling anti-Hebbian and Hebbian (AHaH) computing. This architecture is claimed to integrate the traditionally separate domains of memory and processing, on the way to forming neural networks that can purportedly perform machine-learning tasks at much larger scales and with much greater efficiency than the traditional (von Neumann) computers that most of us are familiar with. If you’d like to read more about this architecture, Knowm’s founders wrote an extremely in-depth paper on the topic. The net result is apparently a computing architecture that’s well-suited to real-time processing for applications like big data analysis.
In addition to its memristors, EE Times says Knowm will offer individually packaged memristors and an emulation suite called Sense so that developers can begin building applications targeted to the platform. We’ll see whether Knowm has more success with its memristor tech than other past efforts have.
Share
