Brainchip OCTOBER 2017 1
Agenda Neuromorphic computing background Akida Neuromorphic System-on-Chip (NSoC) Brainchip OCTOBER 2017 2
Neuromorphic Computing Background Brainchip OCTOBER 2017 3
A Brief History of Neuromorphic Computing Brainchip OCTOBER 2017 4
Semiconductor Compute Architecture Cycles CPU/MPU/GPU Artificial Intelligence Acceleration Disruption Architectural Von Neumann Harvard Multiplicity of ISAs Multiplicity of Vendors Multiplicity of accelerators FPU GPU DSP 1990 AlexNet wins Imagenet Challenge 2012 Acceleration Convolutions Spiking Architecture VLIW Array Memory Datatype Floating Fixed Binary Consolidation 1971 Intel 4004 Introduced X86/RISC GPU FPGA Brainchip OCTOBER 2017 5
The Next Major Semiconductor Disruption $60B opportunity in next decade Training is important, but inference is the major market $M 70,000 60,000 50,000 40,000 30,000 AI Acceleration Chipset Forecast Training Inference General Purpose 20,000 Machine learning requires 10,000 0 dedicated acceleration 2018 2019 2020 2021 2022 2023 2024 2025 Source: Tractica Deep Learning Chipsets, Q2 2018 Brainchip OCTOBER 2017 6
Explosion of AI Acceleration Software Simulation of ANNs Neuromorphic Computing X86 CPU Convolutional Neural Networks X86 CPU Cloud Acceleration Edge Acceleration Re-Purposed Hardware Acceleration Customized Acceleration Google TPU TrueNorth Test Chip Loihi Test Chip Brainchip + Internal ASIC Development OCTOBER 2017 7
Traditional CPU Architecture Inefficient for ANNs Traditional Compute Architecture Artificial Neural Network Architecture Memory Control unit Arithmetic logic unit input output PROCESSOR ACCUMULATOR Optimal for sequential execution Distributed, parallel, feed-forward Brainchip OCTOBER 2017 8
ANN Differences Primary Compute Function Spiking Neural Network Convolutional Neural Network Synapses Reinforced connections Neurons Inhibited connections Spikes Linear Algebra Matrix Multiplication Brainchip OCTOBER 2017 9
Neural Network Comparison Convolutional Neural Networks Spiking Neural Networks Characteristic Result Characteristic Result Computational functions Matrix Multiplication, ReLU, Pooling, FC layers Math intensive, high power, custom acceleration blocks Threshold logic, connection reinforcement Math-light, low power, standard logic Training Backpropagation offchip Requires large prelabeled datasets, long and expensive training periods Feed-Forward, on or off-chip Short training cycles, continuous learning Math intensive cloud compute Low power edge deployments Brainchip OCTOBER 2017 10
Previous Neuromorphic Computing Programs Primarily research programs Investigating neuron simulation 1,000 s of ways to emulate spiking neurons Investigating training methods Academia or government programs SpiNNaker (Human Brain Project) IBM TrueNorth (DARPA) Neurogrid (Stanford) Intel Loihi test chip Brainchip OCTOBER 2017 11
Brainchip OCTOBER 2017 12
Culmination of Decades of Development Brainchip OCTOBER 2017 13
World s first Neuromorphic System on Chip (NSoC) Efficient neuron model Innovative training methodologies Everything required for embedded/edge applications On-chip processor Data->spike conversion Scalable for Server/Cloud Neuromorphic computing for multiple markets Vision systems Cyber security Financial tech Brainchip OCTOBER 2017 14
Akida NSoC Architecture Brainchip OCTOBER 2017 15
Akida Neuron Fabric Most efficient spiking neural network implementation 1.2M Neurons 10B Synapses Able to replicate most CNN functionality Convolution Pooling Fully connected Meets demanding performance criteria 1,100 fps CIFAR-10 82% accuracy Right-Sized for embedded applications 10 classifiers (CIFAR 10) 11 Layers 517K Neurons 616M Synapses Brainchip OCTOBER 2017 16
Neuron and Synapse Counts in the Animal Kingdom Brainchip OCTOBER 2017 17
The Most Efficient Neuromorphic Computing Fabric Relative Implementation Efficiency (Neurons and Synapses) 3X 300X Keys to efficiency Fixed neuron model Right-sized Synapses minimized on-chip RAM 6MB compared to 30-50MB Programmable training and firing thresholds Flexible neural processor cores Highly optimized to perform convolutions Also fully connected, pooling Efficient connectivity Global spike bus connects all neural processors Multi-chip expandable to 1.2 Billion neurons Brainchip OCTOBER 2017 18
Neuromorphic Computing Benefits Top-1 Accuracy 79% Cifar-10 Intel Myriad 2 18 fps/w ~$10 Cifar-10 BrainChip Akida 82% ~$10 83% 1.4K fps/w 80% Cifar-10 IBM TrueNorth 6K fps/w Cifar-10 Xilinx ZC709 6K fps/w ~$1,000 ~$1,000 Tremendous throughput with low power Math-lite, no MACs No DRAM access for weights Comparable accuracy Optimized synapses and neurons ensures precision GoogLeNet Intel Myriad 2 69% ~$10 69% GoogLeNet Tegra TX2 ~$300 4.2 fps/w 15 fps/w Frames per Second/watt Brainchip OCTOBER 2017 19 Note: For comparison purposes only. Data and pricing are estimated and subject to change
Akida NSoC Applications Brainchip OCTOBER 2017 20
Vision Applications: Object Classification Complete embedded solution Flexible for multiple data types <1 Watt On-chip training available for continuous learning Lidar Pixel DVS Ultrasound Sensor Interfaces Conversion Complex SNN Model Object Classification Data Interfaces Neuron Fabric Data Interfaces Brainchip OCTOBER 2017 21
Financial Technology Applications: Fintech Data Analysis CPU Fintech Data Data Interfaces Fintech data distinguishing parameters for stock characteristics and trading information, can be converted to spikes in SW on CPU or by Akida NSoC Conversion Complex SNN Model Pattern Recognition Neuron Fabric Unsupervised learning on chip to detect repeating patterns (Clustering) These trading patterns and clusters can then be analyzed for effectiveness Brainchip OCTOBER 2017 22
Cybersecurity Applications: Malware Detection CPU File or packet properties Data Interfaces Conversion Complex SNN Model File Classification Neuron Fabric Supervised learning for file classification based on file properties File or packet properties distinguishing parameters for files/network traffic, can be converted to spikes in SW on CPU or by Akida NSoC Brainchip OCTOBER 2017 23
Cybersecurity Applications: Anomaly Detection CPU Behavior Properties Data Interfaces Conversion Complex SNN Model Behavior classifiers Neuron Fabric Supervised learning on known good behavior and anomalous behavior Behavior properties can be CPU loads for common applications, network packets, power consumption, fan speed, etc.. Brainchip OCTOBER 2017 24
Creating SNNs: The Akida Development Environment Brainchip OCTOBER 2017 25
AKIDA Training Methods Unsupervised learning from unlabeled data Detection of unknown patterns in data On-chip or off-chip Unsupervised learning with label classification First layers learns unlabeled features, labeled in fully connected layer On-chip or off-chip Brainchip OCTOBER 2017 26
World s first NSoC Low power and footprint of neuromorphic computing Highest performance /w/$ Estimated tape-out 1H2019, samples 2H2019 Complete solution for embedded/edge applications but scalable for cloud/server usage Brainchip OCTOBER 2017 27