SLIDE 1
Brainchip
OCTOBER 2017 |
1
Brainchip
OCTOBER 2017 |
2
1 Brainchip OCTOBER 2017 | Agenda Neuromorphic computing - - PowerPoint PPT Presentation
1 Brainchip OCTOBER 2017 | Agenda Neuromorphic computing - - PowerPoint PPT Presentation
1 Brainchip OCTOBER 2017 | Agenda Neuromorphic computing background Akida Neuromorphic System-on-Chip (NSoC) 2 Brainchip OCTOBER 2017 | Neuromorphic Computing Background 3 Brainchip OCTOBER 2017 | A Brief History of Neuromorphic
SLIDE 2
SLIDE 3
Brainchip
OCTOBER 2017 |
3
Neuromorphic Computing Background
SLIDE 4
Brainchip
OCTOBER 2017 |
4
A Brief History of Neuromorphic Computing
SLIDE 5
Brainchip
OCTOBER 2017 |
5
Semiconductor Compute Architecture Cycles
Disruption Consolidation
CPU/MPU/GPU
- Architectural
- Von Neumann
- Harvard
- Multiplicity of ISAs
- Multiplicity of Vendors
- Multiplicity of accelerators
- FPU
- GPU
- DSP
AlexNet wins Imagenet Challenge
X86/RISC GPU FPGA 1971
Intel 4004 Introduced
Artificial Intelligence Acceleration 2012
- Acceleration
- Convolutions
- Spiking
- Architecture
- VLIW
- Array
- Memory
- Datatype
- Floating
- Fixed
- Binary
1990
SLIDE 6
Brainchip
OCTOBER 2017 |
6
The Next Major Semiconductor Disruption
Source: Tractica Deep Learning Chipsets, Q2 2018
$60B opportunity in next decade Training is important, but inference is the major market Machine learning requires dedicated acceleration
10,000 20,000 30,000 40,000 50,000 60,000 70,000 2018 2019 2020 2021 2022 2023 2024 2025
$M
AI Acceleration Chipset Forecast
Training Inference General Purpose
SLIDE 7
Brainchip
OCTOBER 2017 |
7
Explosion of AI Acceleration
Software Simulation of ANNs
X86 CPU Convolutional Neural Networks Neuromorphic Computing
TrueNorth Test Chip
Customized Acceleration
Edge Acceleration
Re-Purposed Hardware Acceleration
Loihi Test Chip Google TPU
Cloud Acceleration
X86 CPU
+ Internal ASIC Development
SLIDE 8
Brainchip
OCTOBER 2017 |
8 Memory Control unit PROCESSOR Arithmetic logic unit
input
- utput
ACCUMULATOR
Traditional CPU Architecture Inefficient for ANNs
Optimal for sequential execution Distributed, parallel, feed-forward Traditional Compute Architecture Artificial Neural Network Architecture
SLIDE 9
Brainchip
OCTOBER 2017 |
9
ANN Differences – Primary Compute Function
Convolutional Neural Network Spiking Neural Network
Inhibited connections Reinforced connections
∫
Synapses Neurons Spikes
∫
Linear Algebra Matrix Multiplication
∫ ∫
SLIDE 10
Brainchip
OCTOBER 2017 |
10
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 off- chip Requires large pre- labeled datasets, long and expensive training periods Feed-Forward, on or
- ff-chip
Short training cycles, continuous learning
Math intensive cloud compute Low power edge deployments
SLIDE 11
Brainchip
OCTOBER 2017 |
11
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
SLIDE 12
Brainchip
OCTOBER 2017 |
12
SLIDE 13
Brainchip
OCTOBER 2017 |
13
Culmination of Decades of Development
SLIDE 14
Brainchip
OCTOBER 2017 |
14
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
SLIDE 15
Brainchip
OCTOBER 2017 |
15
Akida NSoC Architecture
SLIDE 16
Brainchip
OCTOBER 2017 |
16
Akida Neuron Fabric
Most efficient spiking neural network implementation
1.2M Neurons 10B Synapses
Able to replicate most CNN functionality
Convolution Pooling Fully connected
Right-Sized for embedded applications 10 classifiers (CIFAR 10)
11 Layers 517K Neurons 616M Synapses
Meets demanding performance criteria
1,100 fps CIFAR-10 82% accuracy
SLIDE 17
Brainchip
OCTOBER 2017 |
17
Neuron and Synapse Counts in the Animal Kingdom
SLIDE 18
Brainchip
OCTOBER 2017 |
18
The Most Efficient Neuromorphic Computing Fabric
Relative Implementation Efficiency (Neurons and Synapses)
300X 3X
Fixed neuron model
Right-sized Synapses minimized
- n-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
Keys to efficiency
SLIDE 19
Brainchip
OCTOBER 2017 |
19
Neuromorphic Computing Benefits
Frames per Second/watt Top-1 Accuracy
GoogLeNet Intel Myriad 2
4.2 fps/w 69% ~$10
Cifar-10 Intel Myriad 2
79% 18 fps/w ~$10
Cifar-10 BrainChip Akida
1.4K fps/w 82% ~$10
Cifar-10 IBM TrueNorth
83% 6K fps/w ~$1,000
Cifar-10 Xilinx ZC709
80% 6K fps/w ~$1,000
GoogLeNet Tegra TX2
69% 15 fps/w ~$300
Tremendous throughput with low power
Math-lite, no MACs
No DRAM access for weights Comparable accuracy
Optimized synapses and neurons ensures precision
Note: For comparison purposes only. Data and pricing are estimated and subject to change
SLIDE 20
Brainchip
OCTOBER 2017 |
20
Akida NSoC Applications
SLIDE 21
Brainchip
OCTOBER 2017 |
21
Vision Applications: Object Classification
Lidar Pixel DVS Ultrasound Data Interfaces Neuron Fabric
Metadata Metadata Metadata Metadata
Sensor Interfaces Conversion Complex
01010110 01010110 01010110
SNN Model Object Classification
Data Interfaces
Complete embedded solution Flexible for multiple data types <1 Watt On-chip training available for continuous learning
SLIDE 22
Brainchip
OCTOBER 2017 |
22
Financial Technology Applications: Fintech Data Analysis
Fintech Data Neuron Fabric
Metadata Metadata Metadata Metadata
Conversion Complex
01010110 01010110 01010110
SNN Model Pattern Recognition
Data Interfaces
Unsupervised learning on chip to detect repeating patterns (Clustering) These trading patterns and clusters can then be analyzed for effectiveness
CPU
01010110
Fintech data – distinguishing parameters for stock characteristics and trading information, can be converted to spikes in SW on CPU or by Akida NSoC
SLIDE 23
Brainchip
OCTOBER 2017 |
23
Cybersecurity Applications: Malware Detection
File or packet properties Neuron Fabric
Metadata Metadata Metadata Metadata
Conversion Complex
01010110 01010110 01010110
SNN Model File Classification
Data Interfaces
Supervised learning for file classification based on file properties
CPU
01010110
File or packet properties – distinguishing parameters for files/network traffic, can be converted to spikes in SW on CPU or by Akida NSoC
SLIDE 24
Brainchip
OCTOBER 2017 |
24
Cybersecurity Applications: Anomaly Detection
Behavior Properties Neuron Fabric
Metadata Metadata Metadata Metadata
Conversion Complex
01010110 01010110 01010110
SNN Model Behavior classifiers
Data Interfaces
Supervised learning
- n known good
behavior and anomalous behavior
CPU
01010110
Behavior properties can be CPU loads for common applications, network packets, power consumption, fan speed, etc..
SLIDE 25
Brainchip
OCTOBER 2017 |
25
Creating SNNs: The Akida Development Environment
SLIDE 26
Brainchip
OCTOBER 2017 |
26
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
SLIDE 27
Brainchip
OCTOBER 2017 |
27