异构计算平台神经元/突触规模数据集

本数据集包含神经元/突触规模测试模型代码、神经元规模测试模型的输入输出数据、以及神经形态芯片实测的可配置神经元规模和可配置突触规模数据。项目组面向基于LynMRAM神经形态芯片的大规模异构计算平台开发了神经元/突触规模测试模型，用于测试异构计算平台的神经元规模和突触支持规模。异构计算平台上可配置神经元、突触数量的测试原理是：在模型构建中，设置两个芯片为一组，将该配置文件执行10组，即在20个芯片上执行该配置文件。此外，配置组内第一个芯片的最后一个簇的100个神经元的发放信息传输到第二个芯片的第一个簇，第二个芯片的输入，这样两个芯片形成串序协同工作模式。记录神经网络规模测试模型运行结果并保存结果文件。查看每一拍的神经元脉冲发送结果，每一个bytes记录一个神经元的发送结果，因此结果文件大小（bytes）= 神经元数量（个）；通过突触参数的存储容量对突触数目进行测算，每个突触采用int8的位宽精度进行表达为1字节，因此参数突触存储容量（bytes）= 突触数量（个），运行过程中记录所占用的存储容量。

This dataset contains neuron/synapse scale test model code, input and output data of the neuron scale test model, and configurable neuron scale and configurable synapse scale data measured on neuromorphic chips. The project team developed a neuron/synapse scale test model for the large-scale heterogeneous computing platform based on LynMRAM neuromorphic chips, which is used to test the neuron scale and supported synapse scale of the heterogeneous computing platform. The test principle for configurable numbers of neurons and synapses on the heterogeneous computing platform is as follows: set two chips as a group, run the configuration file for 10 groups, i.e., execute the configuration file on 20 chips in total. In addition, the spike activity information of 100 neurons in the last cluster of the first chip in the configuration group is transmitted to the first cluster of the second chip, which serves as the input of the second chip, thus enabling the two chips to operate in a serial collaborative mode. Record the operating results of the neural network scale test model and save the result files. Check the neuron spike transmission results for each clock cycle: each byte records the transmission result of one neuron, so the size of the result file (in bytes) equals the number of neurons; the number of synapses is calculated based on the storage capacity allocated for synapse parameters. Each synapse is represented with an int8 bit width precision, occupying 1 byte. Thus, the synapse parameter storage capacity (in bytes) equals the number of synapses, and the occupied storage capacity is recorded during the running process.

搜集汇总

背景与挑战

背景概述

该数据集包含用于测试异构计算平台神经元和突触规模的模型代码及相关数据，通过神经形态芯片实测记录可配置的神经元和突触数量。它旨在评估基于LynMRAM芯片的大规模异构计算平台的性能，并基于特定测试原理保存运行结果。