“Enabling Embedded AI for Healthcare,” a Presentation from VeriSilicon
0 likes177 views
Verisilicon, a leader in embedded neural processing units (NPUs), is focused on advancing personal healthcare through AI-enabled devices. The document discusses essential device categories and the challenges for developing non-intrusive, semi-intrusive, and intrusive healthcare devices, emphasizing the need for low-power, efficient systems. Verisilicon aims to provide solutions to these challenges, leveraging their scalable IP portfolio and expertise in creating innovative technology.
“Enabling Embedded AI for Healthcare,” a Presentation from VeriSilicon
- 1. © 2020 VeriSilicon Enabling Embedded AI for Healthcare Shang-Hung Lin VeriSilicon September 2020
- 2. © 2020 VeriSilicon Presentation Agenda • Intro - What do we need? • Device categories for personal healthcare • What is required to make these devices happen? • Device categories - examples • Challenges • Key elements of an Enabling SoC • How VeriSilicon can help? • Conclusion 2
- 3. © 2020 VeriSilicon VeriSilicon – A SiPaaS Company We call it Silicon Platform as a Service, or SiPaaS® • IP-centric • Platform-based • End-to-end turnkey service What we do What we don’t do • No fab • No branded product • No NRE investment • Limited inventory risk 3
- 4. © 2020 VeriSilicon VeriSilicon IP Portfolio – Scalable STAR IPs for Licensing Tablets Smartphone Automotive Wearables & IoT Server Class NPU GPU Video Audio/Voice Display ISP Compression/ Encryption Compute Tablets Computer Vision Imaging DSP FLEXA API 4
- 5. © 2020 VeriSilicon VeriSilicon is The Leader in Embedded NPU IP Since 2016 • Global Leader in Embedded NPU (Neural Processor Unit) • First To Introduce Programmable, Scalable, High Performance, Low Power NPU • First NPU To Support Common Neural Network plus OpenCL • Leading in number of licensed customers (30+) • Leading in number of SOC designs engaged (50+) • Leading in embedded products shipped in mass production • VIP9000 is applied to wide range of applications in AI Vision, AI Voice, and AI Pixels • VIP9000 expands into all major market segments AI VISION AI VOICE AI PIXEL VeriSilicon Vivante NPU Engagements Market Segments & Applications Surveillance IPC Cameras AIOT Smart Home Wearable Automotive Industrial Edge Sever Data Center Mobile Phone 5
- 6. © 2020 VeriSilicon Introduction • Personal healthcare – why? • Comfort, privacy • Customized service, 24-7 care • Great improvement in quality of life • What we need • Advanced AI enabled personal healthcare devices • Sensors, monitors, aides • Sensing/sensor fusion, Interpret, Alarm / Act 6
- 7. © 2020 VeriSilicon What Device Categories Can Help Us? • Non-intrusive • Wearable, body contact, consumer devices • Intrusive • Capsule camera, implants under the skin or in the body • Semi-intrusive • Earbud, hearing aid 7
- 8. © 2020 VeriSilicon What Is Needed To Make These Devices Happen? ►Sense: MEMS sensors ►Interpret: AI engines ►Act: DSP engines ►Communicate: Wireless ►Design: low power / always on 8
- 9. © 2020 VeriSilicon Device Examples Today - Non Intrusive • Fitness Tracker Ring • Wrist band • Chest band 9
- 10. © 2020 VeriSilicon Intrusive Devices For Personal Healthcare • Capsule Camera • Battery operated device; low power inference is crucial • Location detection • Hemorrhage and lesion detection
- 11. © 2020 VeriSilicon Semi-Intrusive • Advanced Hearing aid examples: • In 1994, Hitachi‘s Dr. Makimoto spoke about his vision – a realtime language translator device • In 2005, The Hitchiker‘s Guide to the Galaxy suggested a solution called Babble Fish • At CES 2019, Google assistant does real time translation • Moving forward, we anticipate advanced, AI-powered in-ear devices in market • Active noise treatement • AI-enabled DSP audio signal enhancement • AI-enabled natural user interface 11
- 12. © 2020 VeriSilicon Challenges • Always On – lowest power • Small footprint • Efficient wireless communication • Wireless charging and energy harvesting • Safety and reliability to obtain FDA approval • Security – hacker safe! 12
- 13. © 2020 VeriSilicon Key Elements of An Enabling SoC • Sensor Fusion • Wireless communication • AI engine • DSP engine • Optimized memory subsystem for lowest power operation 13
- 14. © 2020 VeriSilicon AI Functions For True Wireless Earbud Systems (TWS) • Environmental Noise Cancellation (ENC) • Suppress background noise • Hybrid with ANC • Wake Word Trigger • “OK Google”, “Alexa”… • Voice Assistant • Speaker identification • Voice command recognition • Speech recognition, machine translation • Biometric Sensor Fusion • Heart rate, temperature, EKG, VO2… 14
- 15. © 2020 VeriSilicon Optimizing System Design – Example: VeriSilicon PicoAI IP • Tiny AI Solution for Mass Market AIOT devices • Voice and Vision Wake up • AI Voice, AI Vision Applications • 200+ GOPS NPU (VIP9000Pico) • 1 mm2 in TSMC 12nm • Extremely low power Always-on Wakeup capability • Highly integrated, self-contained solution • All Wakeup operations can be achieved without DDR support • Extremely low power design and implementation • Low latency Wakeup • Light weight software stack - RTOS OS, Bare Metal Image sensor MCU/DSP VIP9000- Pico (NPU) MEMORY Wake up Voice sensor Sensor interface PicoAI IP 15
- 16. © 2020 VeriSilicon System Partitioning Is Crucial – Power Hierarchy And Minimizing Memory Accesses 3-Level Waking Up Sensing Processing AI inferencing • Voice Activity Detection (VAD) • ~50 uW • 24 hours/day (Always-On) • Key Word Spotting (KWS) • ~150 uW • ~30 min/day • Voice Command, ASR • 1 mW – 10 mW • ~1 min/day DSP DMAFIFO TCM NPU Local SRAM System Bus DSP DMAFIFO TCM NPU Local SRAM System Bus DSP DMAFIFO TCM NPU Local SRAM System Bus 16
- 17. © 2020 VeriSilicon System Software Design Is Crucial – Software System for Deeply Embedded Systems Network Binary Graph Object Code MCU/DSP VIP9000- Pico MEMORY VIP-Lite Runtime RTOS or Bare Metal Run Time On the TargetOffline Compiling on the Host Tensorflow TF-Lite TF Lite Micro Caffe Darknet Universal Model Converter Post- training Quantizer NN Compiler ACUITY SDK PyTorch Network Binary Graph Object Code 17
- 18. © 2020 VeriSilicon Conclusion • Challenges are ahead • Solutions are starting to form • VeriSilicon is here to help • Let‘s make this Embedded Vision happen together! 18
- 19. © 2020 VeriSilicon Resources Wearable devices: • https://mymotiv.com/the-ring/ • https://www.t3.com/us/features/best-heart-rate-monitor AI-Powered TWS: • https://www.androidcentral.com/best-noise-canceling- true-wireless-earbuds PicoAI, VIP9000, VIP9000Pico: • http://www.verisilicon.com/en/IPPortfolio/VivanteNPUIP ACUITY SDK: • https://github.com/VeriSilicon/acuity-models 2020 Embedded Vision Summit • Visit VeriSilicon’s virtual booth to speak with technology experts and watch exciting demos. 19