"Emerging Processor Architectures for Deep Learning: Options and Trade-offs," a Presentation from Hailo
0 likes1,068 views
The document presents the Hailo-8, a highly efficient edge AI processor with 26 TOPS performance and 3 TOPS/W efficiency. It highlights the processor's flexibility, programmability, and comprehensive SDK that supports various deployment modes. The document also compares its power efficiency and performance against other AI processors in the market.
![© 2019 Hailo 10
Power Efficiency
0
0.5
1
1.5
2
2.5
3
3.5
CLASSIFICATION DETECTION (720p) SEGMENTATION (1080p)
PowerEfficiency[TOP/W/s] NVIDIA AGX XAVIER*
Google Edge TPU*
HAILO Hailo-8
* References:
- NVIDIA AGX Xavier: : https://developer.nvidia.com/embedded/jetson-agx-xavier-dl-inference-benchmarks
- Google Edge TPU: https://coral.withgoogle.com/docs/edgetpu/benchmarks
** Edge TPU performance is measured for 224x224; Linearly extrapolated to 720p
Frame per second (FPS) 656 37 672 8.4* 40 4064](https://image.slidesharecdn.com/e1t4danonhailo2019-190729184921/85/Emerging-Processor-Architectures-for-Deep-Learning-Options-and-Trade-offs-a-Presentation-from-Hailo-10-320.jpg)
"Emerging Processor Architectures for Deep Learning: Options and Trade-offs," a Presentation from Hailo
- 1. © 2019 Hailo Introducing Hailo-8™: The Most Efficient Deep Learning Processor for Edge Devices Orr Danon Hailo May 2019
- 2. © 2019 Hailo Presenting the Hailo-8™ The world’s most powerful and efficient edge AI processor High Performance 26 TOPS High Efficiency 3 TOPS/W Automotive ASIL-B(D) AEC-Q100 Grade 2 Flexible Fully programmable Comprehensive SDK Self Contained No external DRAM Deployment Stand alone mode Co-processor mode
- 3. © 2019 Hailo 3 Power Efficiency 0 1 2 3 4 Hailo-8 Google Edge TPU (*,**) Nvidia Xavier Intel Movidius Myriad X (*) TOPS/Watt Image classification inference task, batch = 1 Based on public benchmarks (see ref below) (*) Excluding host and/or memory (**) Estimated
- 4. © 2019 Hailo Product Overview 4 Hailo Centric System Hailo as Co-Processor
- 5. © 2019 Hailo Current AI Processor Architectures Von-Neumann Architecture Symmetric Dataflow Architecture • Temporal resource allocation • Common memory space • Spatial resource allocation • Segregated memory spaces Fixed Function Accelerator • Theoretically optimal at a specific workload • Minimal flexibility System Bus Control Compute Memory Inter-element Bus Control Compute D-MemI-Mem
- 6. © 2019 Hailo 6 Structure Defined Dataflow Architecture (1) • Model structure defines connectivity • Resources • Heterogeneous • Asymmetric • Variable Control Memory Memory Memory Control Data Interconnect Control Memory Compute Compute ComputeCompute ControlInterconnect
- 7. © 2019 Hailo Structure Defined Dataflow Architecture (2) 7
- 8. © 2019 Hailo Structure Defined Dataflow Architecture (3) 8
- 9. © 2019 Hailo Comprehensive SDK 9 • Full-Stack solution from trained model to deployment • Automatic numerical conversion and emulation • Accurate profiling • Continuous delivery of features and optimizations Numeric Translator Compiler Hailo Devices Emulator Profiler Resource Allocator Model Translator
- 10. © 2019 Hailo 10 Power Efficiency 0 0.5 1 1.5 2 2.5 3 3.5 CLASSIFICATION DETECTION (720p) SEGMENTATION (1080p) PowerEfficiency[TOP/W/s] NVIDIA AGX XAVIER* Google Edge TPU* HAILO Hailo-8 * References: - NVIDIA AGX Xavier: : https://developer.nvidia.com/embedded/jetson-agx-xavier-dl-inference-benchmarks - Google Edge TPU: https://coral.withgoogle.com/docs/edgetpu/benchmarks ** Edge TPU performance is measured for 224x224; Linearly extrapolated to 720p Frame per second (FPS) 656 37 672 8.4* 40 4064
- 11. © 2019 Hailo Hailo-8™ Fast-Track Program • Early access to developer suite for the Hailo-8™ device • Register at our website: hailo.ai 11
- 12. © 2019 Hailo Presenting the Hailo-8™ The world’s most powerful and efficient edge AI processor High Performance 26 TOPS High Efficiency 3 TOPS/W Automotive ASIL-B(D) AEC-Q100 Grade 2 Flexible Fully programmable Comprehensive SDK Self Contained No external DRAM Deployment Stand alone mode Co-processor mode
- 13. © 2019 Hailo Thank You! 13 [email protected]