A brain-inspired computing chip and computing device

Abstract

The invention provides a brain-like computing chip and a computing device. The brain-like computing chip includes a many-core system composed of one or more functional cores, and data transmission is performed between the functional cores through an on-chip network; the functional cores include : at least one neuron processor for calculating multiple neuron models; at least one coprocessor, coupled with the neuron processor, for performing integral operations and / or multiply-add operations; wherein the neuron The metaprocessor may invoke the coprocessor to perform multiply-add-like operations. The brain-like computing chip provided by the invention is used to efficiently implement various neuromorphic algorithms, especially for the computing characteristics of the rich SNN model, synaptic computing with high computing density, and cell body computing with high flexibility requirements.

Description

technical field

[0001] The invention relates to the field of chip technology, in particular to a brain-inspired computing chip and computing equipment. Background technique

[0002] At present, there are many kinds of artificial neuron models according to the degree of sophistication and emphases of simulating biological neurons. Such as the most typical Integrate and fire (IF) model, Leaky integrate-and-fire (LIF) model and Izhikevich model. Traditional brain-like chips (neuromorphic chips) will only consider 1-2 of the models and implement them in a pure ASIC manner. Although most of them have high execution efficiency and low power consumption, they are not programmable enough to support more models . Although a programmable neuromorphic chip with many processor cores is used in traditional chips, each core is a general-purpose ARM processor, but because it does not have an additional acceleration module for multiplying and adding, the calculation efficiency is low. C...

Examples

Embodiment 1

[0054] Embodiment one, Leaky Integrate and Fire (LIF) model

[0055] 1. Update neuron membrane potential value: update the membrane potential value V of LIF pulsed neurons at time t 1 (t), the membrane potential value V of the neuron at time t 1 (t) can be expressed as: axon input X(t) and synaptic weight W 1 The dot product of , and the membrane potential value V maintained by the neuron at the last moment 1rst (t) The sum of the values ​​after the effect of the membrane potential decay coefficient α, as shown in the following formula:

[0056] V 1 (t)=X(t)·W 1 +αV 1rst (t-1) formula (1)

[0057] 2. Neuron output judgment: the output of the LIF pulse neuron is the presence or absence of a pulse represented by 0 or 1, and the output pulse y 1 (t) The neuronal membrane potential V at time t 1 (t) After adding the bias term b (also known as the Changzhi attenuation term), it is related to the neuron threshold V th For comparison, when the biased membrane potential value...

Embodiment 2

[0080] Embodiment 2 Izhikevich Model model

[0081] The Izhikevich Model uses a differential equation to represent the update of the neuron membrane potential, as shown in the following formula, where v represents the membrane potential of the Izhikevich neuron; u is the recovery variable, which is used to describe the Na in the biological neuron + 、K + I represents the input stimulus received by the neuron, similar to the first embodiment of the invention, the input stimulus is obtained by the product of the axon input X and the synaptic weight W.

[0082] Formula (4) shows that the change of neuron membrane potential v over time is determined by itself, the current recovery variable u and the input stimulus I, and formula (5) shows that the change of recovery variable u over time is determined by itself and the current The neuron membrane potential v is jointly determined, and the formula (6) describes the output judgment of the neuron and the reset update logic of the neur...

technical field

[0001] The invention relates to the field of chip technology, in particular to a brain-inspired computing chip and computing equipment. Background technique

[0002] At present, there are many kinds of artificial neuron models according to the degree of sophistication and emphases of simulating biological neurons. Such as the most typical Integrate and fire (IF) model, Leaky integrate-and-fire (LIF) model and Izhikevich model. Traditional brain-like chips (neuromorphic chips) will only consider 1-2 of the models and implement them in a pure ASIC manner. Although most of them have high execution efficiency and low power consumption, they are not programmable enough to support more models . Although a programmable neuromorphic chip with many processor cores is used in traditional chips, each core is a general-purpose ARM processor, but because it does not have an additional acceleration module for multiplying and adding, the calculation efficiency is low. C...

Examples

Embodiment 1

[0054] Embodiment one, Leaky Integrate and Fire (LIF) model

[0055] 1. Update neuron membrane potential value: update the membrane potential value V of LIF pulsed neurons at time t 1 (t), the membrane potential value V of the neuron at time t 1 (t) can be expressed as: axon input X(t) and synaptic weight W 1 The dot product of , and the membrane potential value V maintained by the neuron at the last moment 1rst (t) The sum of the values ​​after the effect of the membrane potential decay coefficient α, as shown in the following formula:

[0056] V 1 (t)=X(t)·W 1 +αV 1rst (t-1) formula (1)

[0057] 2. Neuron output judgment: the output of the LIF pulse neuron is the presence or absence of a pulse represented by 0 or 1, and the output pulse y 1 (t) The neuronal membrane potential V at time t 1 (t) After adding the bias term b (also known as the Changzhi attenuation term), it is related to the neuron threshold V th For comparison, when the biased membrane potential value...

Embodiment 2

[0080] Embodiment 2 Izhikevich Model model

[0081] The Izhikevich Model uses a differential equation to represent the update of the neuron membrane potential, as shown in the following formula, where v represents the membrane potential of the Izhikevich neuron; u is the recovery variable, which is used to describe the Na in the biological neuron + 、K + I represents the input stimulus received by the neuron, similar to the first embodiment of the invention, the input stimulus is obtained by the product of the axon input X and the synaptic weight W.

[0082] Formula (4) shows that the change of neuron membrane potential v over time is determined by itself, the current recovery variable u and the input stimulus I, and formula (5) shows that the change of recovery variable u over time is determined by itself and the current The neuron membrane potential v is jointly determined, and the formula (6) describes the output judgment of the neuron and the reset update logic of the neur...

Images