PORTLAND, Ore.—The first component that realizes the two-decades-old search for analog hardware that mimics the learning ability of the human brain's synapses and neurons has come to market. Based on memristors, the component from startup Knowm Inc. (Sante Fe, N.M.).
According to analysts, Knowm's approach to using unique hardware to process Big Data streams in real time could lead to a unique new model for computer makers worldwide.
"Knowm is interesting for two reasons: its application of a some pretty heavy idea that if executed will have a positive import on the industry trying to move forward with Big Data," service director for technology and software for at Current Analysis (London), Brad Shimmin told EE Times prior to the announcement today. "The biggest challenge of applying machine learning is being able to scale and Knowm's unique perspective may be a able to solve that problem. I expect as they move forward they will need to partner with IBM, Hewlett Packard or Oracle who are trying to solve this problem with Big Data too."
Other analysts agreed, plus added that Knowm's BEOL service will give semiconductor makers worldwide access to Knowm's technology during real-time processing of vast data streams.
"Digital computing infrastructure, based on switching digital bits and separating the functions of persisting data from processing, is now facing some big hurdles with Moore's law. There simply isn't enough power to meet the desires of those wishing to reach biological scale and density, whether evolving artificial intelligence or more practically scaling machine learning to ever larger big data sets," senior analyst and consultant Mike Matchett at the Taneja Group Inc. (Hopkinton, Mass.) told us prior to the announcement today. "Knowm's new solution stack impressively breaks through some of the traditional computing barriers by cleverly leveraging the analog hardware functionality of memristive circuits for both data persistence and direct 'in-memory' computing. They are providing a new way to approach machine learning at scale with chips of their new 'synapses', larger scale simulators, development libraries, and services to help layer on their memristive based designs right on top of traditional CMOS designs (through Back End of Line layering)."
Knowm's unique perspective comes from its combo 'Anti-Hebbian and Hebbian' (AHaH) machine learning approach using memristors (others use one or the other, but only Knowm uses both at once) thus allowing the customer to define their own specific learning algorithm using the same building blocks.
"Knowm’s AHaH computing approach combines the best of machine learning and quantum computing via memristors," Chief Executive Officer Alex Nugent told EE Times in advance of the company's unveiling today. "Our neuromemristive processors use a low-level instruction set that can be combined in various ways to achieve any number of learning algorithms."
Each memristor 'remembers' how much current has passed through it, and in what direction, by changing its resistance, here based on mobile metal ion conduction through the chalcogenide material.
Unlike quantum computing, which needs to establish a delicate qubit and design a novel — hard to manufacture — non-destructive read mechanism, "AHaH Computing does not rely on the assumption of a 'non-destructive read'. Every access is assumed to result in adaptation. By understanding the resulting attractor states in AHaH Computing, its possible to design systems that continuously adapt and learn."
Instead of configuring its memristors into a crossbar array — the bane of Hewlett Packard and Hynix and their unproductive foray (so far) into memristor-based supercomputers, Knowm configures them into kT (Boltzman's constant, k, times temperature, T) random access memory (RAM), or kt-RAM cells, that contain differential-output memristor arrays with an S-RAM cell to connect/disconnect them from a H-shaped fractal interconnect. The H-interconnect can be tiled into arrays of any size, and will be offered to semiconductor makers as a top-layer added in Knowm's own fab to their complementary metal oxide semiconductor (CMOS) application specific integrated circuits (ASICs).
Packaged memristors will also be offered to customers in 16-pin DIP packages including eight memristors, for testing, along with a complete emulator--called Sense--for prototyping deep learning algorithms to be executed by the kT-RAM arrays that Knowm fabricates atop their customer's own CMOS ASIC. The Sense emulator software is based on a Knowm application programmer's interface (API) that runs in Java for evaluation of machine learning applications. To prove the concept, Knowm supplies an example application for anomaly detection using spikes, built on top of existing open source programs Bro, Logstash, Elastic Search, KiIbana and Apache Storm. Input from those programs goes into Knowm's Anomaly which detects and identifies anomalies in any data stream.
The Knowm memristor chip is designed for engineers to verify that the actual resistance changes are accurately mimicked by the emulator (see next figure).
Knowm's memristors are not based on the technique of oxygen vacancy migration as is HP's, but on metal ion migration developed by professor Kris Campbell at Boise State University (Idaho). The multi-layer material, however, achieves the same results as HP's--changing the resistance of the material up-or-down, depending on how much voltage is applied to it and in which direction.
Knowm Web Application Server allows customers to develop and test Epiphany applications for persistent real-time streaming data anomaly detection under Apache Storm on four parallel hosts optimized for high synaptic bandwidth of the kT-RAM emulator.
"Our devices are fabricated with a layer of metal that is easily oxidizable (typically either Ag or Cu), located near one electrode of the two-electrode device. When a voltage is applied across the device with the more positive potential on the electrode near this metal layer, this metal is oxidized to form silver or copper ions. Once formed, these metal ions move through the device towards the other, lower potential, electrode. The metal ions move through a layer of amorphous chalcogenide material (the active layer) to reach the lower potential electrode where they are reduced to their metallic form and eventually form a conductive pathway between both electrodes that spans through the active material layer, lowering the device resistance. Reversing the direction of the applied potential causes the conductive channel to dissolve and the device resistance to increase. The device is thus a bipolar device, cycling between high and low resistance values by switching the polarity of the applied potential," Campbell told us.
The inventor of the memristor name, professor Leon Chua at the University of California, Berekeley, is still backing the oxygen vacancy migration method of building memristors, because they can be made as small as eight-nanometers, whereas Campbell's smallest implementation so far is 100 nanometers, but he nevertheless endorsed Knowm.
"The nice thing about Dr. Campbell's memristor is that it is the first memristor device available commercially. This will help university professors to start offering laboratory experiments for students. I was so impressed by her device that I have invited Kris to present a live demonstration at my keynote opening lecture at the 30th anniversary of the 2015 Midwest Symposium on Circuits and Systems on Aug. 2 at Fort Collins, Colorado," Chua told EE Times in advance of the announcement today.
For production runs, Knowm's own fab will fashion the user's emulated memristor array atop the CMOS chip supplied to them by the customer.
Perhaps Knowm's business model is its most unique aspect. The company currently runs without venture capital, but instead from angle-investors, primarily chief administrative officer (CAO) Hillary Riggs and consultant Sam Barakat. Its revenue streams will come from its memristor chips, its kt-RAM BEOLn services, its development systems which it hopes to have running on all available platforms (not just its own server) and royalties it pays to application developers. In other words, customers that design applications that Knowm sells will share in that revenue stream.
— R. Colin Johnson, Advanced Technology Editor, EE Times
Join over 2,000 technical professionals and embedded systems hardware, software, and firmware developers at ESC Silicon Valley July 20-22, 2015 and learn about the latest techniques and tips for reducing time, cost, and complexity in the embedded development process.
Passes for the ESC Silicon Valley 2015 Technical Conference are available at the conference’s official site with discounted advance pricing until July 17, 2015. The Embedded Systems Conference and EE Times are owned by UBM Canon.