Brainchip Holdings Ltd. · ASX:BRN
Brainchip — Investor Presentation (September 2018)
27 slides from Brainchip's investor presentation, Sep 10, 2018
Slide 1: Brainchip
For personal use only Brainchip O CTO BER 2017 | 1
Slide 2: Agenda
Agenda For personal use only Neuromorphic computing background Akida Neuromorphic System-on-Chip (NSoC) Brainchip O CTO BER 2017 | 2
Slide 3: Neuromorphic Computing Background
For personal use only Neuromorphic Computing Background Brainchip O CTO BER 2017 | 3
Slide 4: A Brief History of Neuromorphic Computing
A Brief History of Neuromorphic Computing For personal use only Brainchip O CTO BER 2017 | 4
Slide 5: Semiconductor Compute Architecture Cycles
Semiconductor Compute Architecture Cycles For personal use only CPU/MPU/GPU Artificial Intelligence Acceleration Acceleration Architectural Convolutions Von Neumann Spiking Architecture Disruption Harvard VLIW Multiplicity of ISAs Array Multiplicity of Vendors Memory Multip
Slide 6: The Next Major Semiconductor Disruption
The Next Major Semiconductor Disruption For personal use only $60B opportunity in next decade AI Acceleration Chipset Forecast 70,000 60,000 Training Training is important, but 50,000 Inference inference is the major 40,000 General Purpose market $M 30,000 20,000 Machine learning requires 10,000 ded
Slide 7: Explosion of AI Acceleration
Explosion of AI Acceleration For personal use only Neuromorphic Computing Convolutional Neural Networks Software Simulation of ANNs X86 CPU X86 CPU Cloud Acceleration Edge Acceleration Re-Purposed Hardware Acceleration Google TPU Customized Acceleration TrueNorth Loihi Test Chip Test Chip Brainchip
Slide 8: Traditional CPU Architecture Inefficient for ANNs
Traditional CPU Architecture Inefficient for ANNs For personal use only Traditional Compute Architecture Artificial Neural Network Architecture Memory Control Arithmetic input unit logic unit output PROCESSOR ACCUMULATOR Optimal for sequential execution Distributed, parallel, feed-forward Brainchip
Slide 9: ANN Differences – Primary Compute Function
ANN Differences – Primary Compute Function For personal use only Spiking Neural Network Convolutional Neural Network ∫ ∫ ∫ Synapses Reinforced connections Linear Algebra Matrix Multiplication ∫ Neurons Inhibited connections Spikes Brainchip O CTO BER 2017 | 9
Slide 10: Neural Network Comparison
Neural Network Comparison For personal use only Convolutional Neural Networks Spiking Neural Networks Characteristic Result Characteristic Result Computational Matrix Multiplication, Math intensive, high Threshold logic, Math-light, low functions ReLU, Pooling, FC power, custom connection power, sta
Slide 11: Previous Neuromorphic Computing Programs
Previous Neuromorphic Computing Programs For personal use only 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 (Stanfo
Slide 12: For personal use only
For personal use only Brainchip O CTO BER 2017 | 12
Slide 13: Culmination of Decades of Development
Culmination of Decades of Development For personal use only Brainchip O CTO BER 2017 | 13
Slide 14: World's first Neuromorphic System on Chip (NSoC)
For personal use only 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 sys
Slide 15: Akida NSoC Architecture
Akida NSoC Architecture For personal use only Brainchip O CTO BER 2017 | 15
Slide 16: Akida Neuron Fabric
Akida Neuron Fabric For personal use only Most efficient spiking neural Right-Sized for embedded network implementation 1.2M Neurons applications 10B Synapses 10 classifiers (CIFAR 10) Able to replicate most CNN 11 Layers functionality 517K Neurons Convolution Pooling 616M Synapses Fully connected M
Slide 17: Neuron and Synapse Counts in the Animal Kingdom
Neuron and Synapse Counts in the Animal Kingdom For personal use only Brainchip O CTO BER 2017 | 17
Slide 18: The Most Efficient Neuromorphic Computing Fabric
The Most Efficient Neuromorphic Computing Fabric For personal use only Relative Implementation Efficiency Keys to efficiency (Neurons and Synapses) Fixed neuron model Right-sized Synapses minimized 300X on-chip RAM 6MB compared to 30-50MB Programmable training and firing thresholds Flexible neural p
Slide 19: Neuromorphic Computing Benefits
Neuromorphic Computing Benefits use only 83% Cifar-10 IBM ~$1,000 Cifar-10 TrueNorth Tremendous 82% ~$10 BrainChip throughput with low Top-1 Accuracy Akida 6K fps/w power For personal 1.4K fps/w Math-lite, no MACs Cifar-10 Cifar-10 No DRAM access 79% ~$10 80% ~$1,000 Intel Myriad 2 Xilinx ZC709 for
Slide 20: Akida NSoC Applications
Akida NSoC Applications For personal use only Brainchip O CTO BER 2017 | 20
Slide 21: Vision Applications: Object Classification
Vision Applications: Object Classification For personal use only Complete embedded solution SNN Model Object Classification Flexible for multiple data types Lidar Conversion Complex Sensor Interfaces 01010110 Data Interfaces Metadata Pixel 01010110 <1 Watt 01010110 Metadata DVS Metadata On-chip trai
Slide 22: Financial Technology Applications: Fintech Data Analysis
Financial Technology Applications: Fintech Data Analysis For personal use only SNN Model Fintech Data Pattern Recognition Unsupervised Conversion 01010110 learning on chip to 01010110 Complex 01010110 01010110 detect repeating Data Interfaces CPU patterns (Clustering) These trading Neuron Fabric pat
Slide 23: Cybersecurity Applications: Malware Detection
Cybersecurity Applications: Malware Detection For personal use only File or packet SNN Model properties File Classification Supervised Conversion 01010110 learning for file 01010110 Complex 01010110 01010110 CPU classification Data Interfaces based on file properties Metadata Metadata Neuron Fabric
Slide 24: Cybersecurity Applications: Anomaly Detection
Cybersecurity Applications: Anomaly Detection For personal use only Behavior SNN Model Properties Behavior classifiers Supervised learning Conversion 01010110 on known good 01010110 Complex 01010110 01010110 CPU behavior and Data Interfaces anomalous behavior Metadata Metadata Neuron Fabric Metadata
Slide 25: Creating SNNs: The Akida Development Environment
Creating SNNs: The Akida Development Environment For personal use only Brainchip O CTO BER 2017 | 25
Slide 26: AKIDA Training Methods
AKIDA Training Methods For personal use only 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 O CT
Slide 27: World's first NSoC
For personal use only 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 O CTO BER 2017 | 27
BrainChip Holdings — Investor Presentation (September 2018) on 6ix