Brainchip Holdings Ltd. · ASX:BRN
Brainchip — Linley Conference 2019
15 slides from Brainchip's conference presentation, Oct 24, 2019
Slide 1: Introducing Akida
Introducing Akida NEURAL PROCESSING UNIT Anil Mankar Chief Development Officer & Co-Founder Linley Conference | October 24, 2019 BRAINCHIP CONFIDENTIAL
Slide 2: About BrainChip
About BrainChip BrainChip is a global technology company that has developed Akida, a revolutionary advanced neural networking processor that brings artificial intelligence to the edge in a way that existing technologies are not capable. The solution is high-performance, small, ultra-low power and en
Slide 3: Akida Platform
Akida NPU EDGE AI Technology Object Detection/classification Keyword Cyber- Data analytics MobileNet/CFAR10 Etc Vibration other Security Akida Platform Vision Internet Time series Lidar 3D Audio Other Sensor Camera frames Packets multivariate Point clouds Data Data or DVS Events Data Data Data to ev
Slide 4: Why Use Spiking Event Based Processing
Why Use Spiking Event Based Processing Spiking Events are non-zero activations We only process Events Events are inherently sparse leading to reduced number of operations Quantizing weights and activations to 1,2 or 4 bits reduces memory requirements Each Layer computations are done on allocated NPU
Slide 5: Architectural Differentiation
Akida Architectural Differentiation Standard Deep Learning Accelerators (DLA) Perform a MAC function on arrays of numbers. The neural network is executed on the host CPU. Intermediate results are stored in host CPU memory. Akida Event Based Neural network Processing Runs the entire network on NPU co
Slide 6: NSoC: A Perfect solution for Edge AI Devices
Akida NSoC: A Perfect solution for Edge AI Devices Single low-power, hardware platform to run: Conventional CNN/DNN inference algorithms On-chip incremental learning at EDGE devices Native SNNs with event-based learning algorithm Entire Multi-layer Neural Network runs on AKIDA Fabric Preprocessed da
Slide 7: NSoC Chip: Complete AI Edge Solution
Akida NSoC Chip: Complete AI Edge Solution Single Platform for CNN inference or Native SNN with on chip learning for EDGE Devices Data input Interfaces On-Chip Processor ⁃ PCI-Express 2.1 x2 Lane Endpoint ⁃ M-Class CPU with FPU & DSP ⁃ USB 3.0 Endpoint ⁃ System management ⁃ I3C, I2S, UART, JTAG ⁃ Ak
Slide 8: Configurable Intellectual Property Blocks
Akida : Configurable Intellectual Property Blocks : Configurable number of NPUs Programmable Data to Spike Converter Event Based Processing Engines running on a single clock Configurable on-chip SRAM memory Runs full Native SNN in Hardware Runs full converted CNN/DNNs in Hardware On chip communicati
Slide 9: Akida Chip Simulator
Akida CHIP SIMULATOR 10/27/2019 BRAINCHIP CONFIDENTIAL | INTRODUCING AKIDA | 9
Slide 10: Akida CNN2SNN Flow
Akida CNN2SNN FLOW 10/27/2019 BRAINCHIP CONFIDENTIAL | INTRODUCING AKIDA | 10
Slide 11: Mapping CNN to SNN for Inference
Akida : Mapping CNN to SNN for Inference MobileNet V1 30 layers with many MAC operations Entire network mapped onto separate Akida NPUs (4,208,224 parameters ) All layers run in parallel storing results within the distributed SRAM memory on chip Performance from the chip simulator Speed: 80 FPS Powe
Slide 12: Edge Inference Simulated Demo
Akida : Edge Inference Simulated Demo MobileNet and VGG-16 ported to AKIDA What does this mean? Any visual classification task that is done by a MobileNet or VGG DNN can be ported to AKIDA 10/27/2019 BRAINCHIP CONFIDENTIAL | INTRODUCING AKIDA | 12
Slide 13: Native learning on-chip
Native learning on-chip ON-CHIP LEARNING APPLICATION USING DVS CAMERA Akida trained on hand-gestures using event-based camera (DVS) Real time learning and detection of patterns Uses our proprietary unsupervised learning rule layer filters size input output (pre- processing) 128x128x1 64x64x1 conv 4
Slide 14: Akida Native SNN Mode
Akida Native SNN Mode Example network is a cyber-security classification task using the CSE-IDS-2018 database (220.8 Gb) 15 classes (including ‘normal’) Training time is 2.2 hours, single epoch, including preprocessing on a simple i7 laptop with no GPU Automatic labeling of active neurons extracted
Slide 15: Summary - A Holistic Approach to Edge AI
Akida Summary - A Holistic Approach to Edge AI Changing the game for Edge AI Run Full network on the HW efficiently Game changing FPS/W Parallel execution , lower power and real time performance Lower minimum power Avoid external memory accesses , reduce bit computation Fewer TOPS required for deliv