About the AltAI-1 neural processing unit
The AltAI-1 neural processing unit is a very large-scale integration (VLSI) circuit with a computational architecture that imitates the operation of biological sensory and nervous systems. The AltAI-1 neuromorphic processor is intended for the execution (inference) of spiking neural networks created using platform components or using ternary layers of the ANN2SNN package. The AltAI-1 architecture implements the "computation next to memory" principle, which makes it possible to avoid excessive power consumption (typical of the von Neumann architecture) for transferring data between the memory block and the computational core. The AltAI-1 neural processing unit has a low power consumption.
To execute a spiking neural network, connect the backplane with the AltAI-1 neuromorphic processing unit to the computer via a PCI-E or USB interface.
The elementary unit of an impulse neural network is a neuron, which has a set of input signal lines (dendrites) and an output signal line (axon). The state of a neuron is described by neural potential. If, during the execution of the impulse neural network, the neuron potential exceeds the specified threshold, the neuron emits a spike along its axon and sends it to the specified target neurons. A spike can be either positive or negative. When modeling a neuron, the leak mechanism is also taken into account; the leak mechanism is intended for the relaxation of the neural potential.
The neurons of the impulse neural network are connected by synapses. Each synapse can be assigned a weight by which the neuron potential changes when a spike is received from a neuron associated with this synapse.
The main structural unit of the AltAI-1 neural processing unit architecture is a core. The core combines a group of neurons and memory for storing their parameters. The function of the neurons in the core is performed by a state machine that models the neuron behavior.
All neurons in a spiking neural network are synchronized with the help of a tick signal. After a signal is given, potentials of all neurons are sequentially updated by all cores, and spikes are emitted if needed. During a tick, the core performs 262,144 synaptic operations. The tick frequency does not exceed 2,000 Hz. One synaptic operation is performed in one clock cycle with a frequency of 600 MHz.
The AltAI-1 neuromorphic processor has a regular structure of cores in the form of a rectangular network, where each core is directly connected to its 4 neighbors. To transfer spikes between the neurons of different cores, a mechanism is implemented for routing such signals through the core network. Accordingly, when the function of a neuron is performed, the values of the neural potentials sent by the neurons from other cores can be added to the potential of the modeled neuron.
The AltAI-1 neuromorphic processor provides the ability to scale by increasing the number of VLSI chips. For questions regarding scaling, please contact Kaspersky specialists.
The table below provides general information about the AltAI-1 neuromorphic processor and its components:
General information about the AltAI-1 processor
Property |
Restriction |
|
|---|---|---|
VLSI |
Crystal area |
6 sq. mm |
Manufacturing technology |
28 nm |
|
Maximum number of neuromorphic cores |
256 |
|
Number of neuromorphic cores in the prototype |
16 |
|
Maximum number of modeled neurons |
131 072 |
|
Number of modeled neurons in the prototype |
8192 |
|
Core
|
Number of modeled neurons |
512 |
Number of input lines |
512 |
|
Range of neuron potential values |
–32,768 to 32,767 The neuron potential is stored as a 16-bit value. |
|
Range of synapse weight values |
The synapse weight is stored as an 8-bit unsigned value. |
|
Variation of weights per neuron |
4 |
|
Number of synaptic operations |
9 Each synaptic connection between the neuron being modeled and the input line is characterized by one of the nine synaptic operations. You can control the impact of this synaptic connection on the neuron by selecting an operation. The following synaptic operations are available:
|
|
Number of synaptic operations per tick |
262 144 |
|
Tick frequency |
No more than 2,000 Hz |
|
Core power consumption |
No more than 4 mW |
|
Core memory size |
88 KB |
|
Core memory size for storing neuron parameters |
16 KB |
|