NeuroMem for real-time machine learning and practical artificial intelligence
NeuroMem® is a unique architecture of Neuromorphic Memories which react to digital stimuli and can learn and recognize in real-time. A bank of NeuroMem chips can be compared to a brain because of its low power requirement, built-in adaptive learning and interconnectivity to make global decision. Stimuli can derive from any data types such as text, scientific datasets, bio-signals, audio files, images and videos, etc.
Learn on the chip, or load models previously learned and saved by the same or other chip
Recognize patterns in micro-seconds regardless of the number of models stored in the neurons.
100,000 recognitions per second for milliWatts per 1,000 models stored in the neurons
Available on ASICs, SOCs, and FPGAs and can be evaluated on a variety of platforms
The Flavors of NeuroMem…
The NeuroMem technology is presently deployed in two commercial silicon chips and available as IP core for FPGA and SOCs
- A pure neuronal chip , cascadable
- 1024 neurons
- 256 bytes of memory per neuron
- Sensory digital input bus and recognition stage
- Evaluation platforms: Braincard module, NeuroStack board
Intel Quark SE
- A microcontroller with pattern recognition accelerator
- 128 neurons
- 128 bytes of memory per neuron
- Evaluation platforms: Curie module, Arduino/Genuino 101 board
The General Vision CM1K is a 95% neuron-only chip. It also features a build-in recognition stage accepting digital input lines from CMOS and other sensors. CM1K is cascadable, so multiple chips can be connected to build networks with a capacity of one thousand, tens of thousand and even millions of neurons.
The Intel Quark SE is a microcontroller with an onboard sensor subsystem and neurons designed for wearable devices and consumer and industrial edge products.. The neurons are used as a pattern matching accelerator to learn and differentiate patterns from input signals. The Intel Curie module is the assembly of a Quark SE with a 6-axis combo sensor and flash memory.
The NeuroMem technology is available for licensing as an IP core for integration into Field Programmable Gate Arrays or System On a Chip. More…
For more information please inquire with us.
Silicon Neurons Inspired by Biology
- Neurons are all interconnected and working in parallel, recognizing or learning one pattern in a constant number of nsec and this regardless of the number of neurons committed in the chip
- Neurons can behave collectively as (1) a K-Nearest Neighbor or (2) a Radial Basis Function (more specifically a Restricted Coulomb Energy classifier). They can cope with ill-defined and fuzzy data, high variability of context and novelty detection.
- Neurons feature a collective built-in model generator which means that learning is done on the chip!
- Last, but not least, multiple NeuroMem chips can be daisy-chained to scale a network from thousands to millions of neurons with the same speed performance and simplicity of operation as a single chip.
State of the Art for neuromorphic chips
In the past decade, the renewed interest for neuromorphic chips has awakened competition from companies such as IBM with its TrueNorth chip, QualComm with the Zeroth chip, and more (see market report). Their periodic mediatic campaigns are getting everyone accustomed to the fact that brain-like chips are coming to the market, but where are these chips? And to whom will they be accessible?
- The ZISC chip was invented by Guy Paillet, our CEO, and jointly developed with IBM-France in 1993, at the same time as the joint venture between Nestor and Intel was working on the NI1000 chip. The CM1K survived and inherited a successor in 2007 with the CM1K chip from General Vision.
- Prestigious universities and laboratories are designing neuromorphic chips such as the MIT, Stanford, GorgiaTech, and more, but again when will they be available to the public? how easy to use?
- It seems that today, only the NeuroMem neurons are commercially available and users have 2 choices: Our CM1K chip with 1024 neurons (expandable and with full access to the neurons), or the Curie module from Intel with 128 neurons and limited access to the neurons.
- What about FPGA? The NeuroMem IP is available for FPGA and especially suitable for SOC running at high frequency. However, it is a viable solution only up to a certain number of neurons, after which the size and cost of the FPGA become unpractical for consumer appliances and IoT.