Related Documents
The CM IP package enables a thorough evaluation of our IP blocks prior to their integration into an ASIC or SOC. It is composed of an evaluation board featuring an FPGA in which the selected IP cores can be instantiated as well as a connector to support the addition of a CMEnK neuron expansion module. The IP cores are delivered as black boxes.
| DESCRIPTION | PRICE |
BUY |
|---|---|---|
| CogniMem™ IP Evaluation Kit |  |
The CM IP Kit includes a CogniBlox Module with its FPGA already programmed with one instantiation of 32 neurons already daisy-chained with the 4 CogniMem CM1K Chips of the board. This platform allows testing the full compatibility between the soft IP and hardware IP by combining both in a same neural network. The FPGA is ideal to design your interface between input data sources and the neurons.
This platform is ideal to design and evaluate your own recognition engine based on the CogniMem parallel neural network. Refer to the CM1K and CogniBlox datasheet for more information.
FEATURES
- Recognize one vector among any number, in 10 microseconds with a 27 Mhz clock
- Recognition time independent from the number of models
- Learn a vector in 10 microseconds
- Save and load models (i.e. your knowledge base)
- Simple RTL instructions (less than a dozen registers)
- Parallel and serial communication
A CogniMem neural network has a very simple architecture: it is a chain of identical neurons operating in parallel. A neuron is an associative memory which can autonomously compare an incoming pattern with its reference pattern. During the recognition of an input vector, all the neurons communicate briefly with one another (for 16 clock cycles) to find which one has the best match. Optionally the CM-IP can integrate a built-in recognition engine which can receive vector data directly through a digital input bus, broadcast it to the neurons and return the best-fit category 3 microseconds later. In the case of a video input signal, CogniMem can optionally extract a 1D vector from 2D video data.
PRODUCT SPECIFICATIONS
- Patented parallel architecture with 1024 neurons
- RCE (Restricted Coulomb Energy) neural network
- Radial Basis Function (RBF) or K-Nearest Neighbor (KNN) classifier
- Vector data: up to 256 bytes
- Classification status: Identified, Uncertain or Unknown
- Categories: up to 32768 values
- Distance calculation: L1 or LSup distance norms
- Sub-networks: up to 127 context values
- Trained by example
- Active learning, supervised and unsupervised