Slide 1: Snapdragon 865 5G
1024x2048 @ 31.25 FPS
Slide 2: Qualcomm AI Innovation Challenge
Qualcomm Al > Innovation Challenee : TensorFlow OPPO Testinai Lite mp me wk SB HK 2 Ivac *® CETZ DES6ERRRE EXTREME VISION Raa CSDN) fhunmders=ort cYZONE
Slide 3: Hero Apps and Developers
> NS aera l ees i Video placeholder Qualcomm Al Innovation Challenge a 20 Al Innovation Challenge + Hero Apps 12,000+ Developers Touchbased
Slide 4: Hardware
Hardware
Slide 5: AI Hardware
Video placeholder Al Hardware et eas i ‘no CCCCCOOEEL CCC EET TCC COECCCC Ef
Slide 6: Qualcomm Snapdragon product notice
Qualcomm Snopdragon is a product of Qualcomm Technologies, Inc. and/or its subsidiaries.
Slide 7: Generation: Qualcomm AI Engine
th Generation Qualcomm Al Engine
Slide 8: Qualcomm AI Engine
Hexagon 780 Processor Qualcomm Hexagon” 780 Processor Qualcomn Qualcomm Qualcomm Sensing Processor Kryo Hub Security 680 CPU Qualcomm” : Snapdragon* X60 SG Modem-R
Slide 9: Fused AI accelerator
Fused Al accelerator C 6 a 6 i £9 a 00 09 oN 5 G Qualcomm Kryo 680 CPU Processor Qualcomm Securi Qualcomm Sensing Hub Qualcomm Snapdragon™ — X60 5G Modem-R
Slide 10: Slide 10
Slide 11: Adreno GPU
== Adreno 650 GPU Vector Tensor Scalar Processor Kryo Security 585 CPU
Slide 12: Hexagon Processor
Hexagon 700 Processor Tensor * 16X dedicated memory for single application + Up to 1000X hand off time improvement in certain use cases + Seamless workload sharing between Scalar, Vector and Tensor Vector eXtensions Tensor Accelerator for matrix Scalar Accelerators 50% . performarked Processing. imp
Slide 13: Qualcomm Adreno GPU
Qualcomm’ Adreno” 660 GPU Tensor Vector = Qualcomm’ > Sensing
Slide 14: Fused AI Accelerator
Fused Al Accelerator
Slide 15: Trillion operations per second
Trillion operations per second Snapdragon 888 Snapdragon 865
Slide 16: Peak performance on classification networks
Peak performance on classification networks Inference: (inf/s) 1110 600
Slide 17: Power consumption
Power consumption Efficiency: (inf/W) Snapdragon 888 150 TD «0 Resnet 50 310 MobileNet V2 120 InceptionV3 JJ)
Slide 18: Slide 18
Slide 19: 2nd Gen Qualcomm Sensing Hub
2™4 Gen Qualcomm’ Sensing Hub Qualcomm’ Hexagon” 780 Processor Qualcomm’ Sensing
Slide 20: Always-on low-power dedicated hardware
Always-on low-power dedicated hardware Al processor Qualcomm‘ Sensing g Hub Qualcomm’ Hexagon” 780 Processor Processor Kry
Slide 21: AI performance
Al performance With 274 Gen Qualcomm Sensing Hub Ist G 5X Task Reduction is! en L : Qualcomm Sensing Hub > ' 1 offload from Hexagon Processor: 3ased on INT8, GOPS/s, LPI performance
Slide 22: Ambient audio and event detection
Ambient audio / audio event detection Screen wake Lift detection Car crash detection Qualcomm Sensing Hub TensorFlow Micro Activity recognition Earthquake detection
Slide 23: Hot word detection on Snapdragon using TensorFlow Micro
Hot word detection on Snapdragon 888 using TensorFlow Micro 2nd Gen 3 8 Qualcomm % Sensing Hub Task offload from Hexagon processor TensorFlow Micro
Slide 24: Video placeholder
Video placeholder a
Slide 25: 5G Camera
5G Camera Video placeholder Al Hardware oo Sounds "Chaotic" increasing ring volume so you don't miss a call = + a Audio Analytic i ek: a “= Sounds "Chaotic" @ Sounds “Lively” Sounds "Calm"
Slide 26: Snapdragon and TrinamiX
Video placeholder Al Hardware Qualcomm 'g) snapdragon trinamix A brand of BASF — We create chemistry
Slide 27: Slide 27
Slide 28: Qualcomm Neural Processing SDK
Qualcomm’ Neural Processing SDK Ever Powering over on-device Al SDK Android Devices
Slide 29: Hexagon NN direct
SS Hexagon NN direct on Qualcomm*® Snapdragon” 865 aoets wettegentee ) memoly Qualcomm Adreno™ 650 GPU
Slide 30: Qualcomm and Android Neural Networks API
f A Qualcomm & Android Neural season Neural Processing SDK ti Networks API nai Qualcomm” wisp Al Engine direct Qualcomn” Acireno 660 GPU Qualcomm’ Hexagon 780 Processor Qualcomm" Kryo" 680 CPU
Slide 31: Unified AI API
| 4 UL LL. d vv fs in Qualcomm* Biww Al Engine direct Accessibility for everyone Unified Al API available across Qualcomm Al Engine (Hexagon, CPU, GPU) Compatible with 5th gen Qualcomm Al Engine Modularity & Extensibility Per accelerators & operation
Slide 32: Hugging Face on Snapdragon
e Video placeholder Hugging Face ® A is Ks ra $ Qualcomm >) * sngpdragon 5S i> 4 4
Slide 33: User defined operators
User defined operators operators Hexagon OpenCL SDK
Slide 34: quantized_add_generic
quantized_add_generic(size, a, b, a_offset, b_offset, a_mult, b_mult, output, target, ctx): A_offset = tvm.var('A_offset', dtype='uint8') B_offset = tvm.var("B_offset', dtype="uints') A_mult = tvm.var(‘A_mult', dtype="uinti6' B_mult = tvm.var('B_mult', dtype="uinti6') N = tvm.var("N") A = tvm.placeh
Slide 35: Years
years research product
Slide 36: Years of research to product
13 years of research to product @ Research face @ MWC demo. © Collaboration @ Brain Corp. ® Qualcomm © Power efficiency © Mobile Al detection with showcasing with Google on raises $114M Atificial Intelligence | gains through Enablement Deep-learning deep learning photo sorting TensotFlow Research in
Slide 37: Slide 37
Slide 38: Snapdragon
S G snapdragon
Slide 39: AI Model Efficiency Toolkit
Qualcomm Neural Processing SDK & Al Dedicated Al Model Efficiency Toolkit accelerator New features and First to support improvements Tensorflow Micro = Highlights Qualcomm Al | Engine direct | Easier and faster access Fused Al Accelerators to the entire Al Engine * Tensor - 2X compute capacity * Sca
Slide 40: Thank you
Qualcomm Thank you Follow us on: f W in For more information, visit us at: www.qualcomm.com & www.qualcomm.com/blog Nothing in these materials is an offer to sell any of the References in this presentation to “Qualcomm” may mean Qualcomm components or devices referenced herein, Incorporated, Qualcom