OpenTalks.AI - Илья Левчук, ИИ для телеком сектора в новом цифровом мире

Transcript

1. AI for the Telecom sector in the new digital world Ilya Levchuk CNBG, AI Taskforce
2. Huawei Confidential 2 AI: Mostly in cloud, some at the edge 10 changes that shape the future & Huawei AI strategy Today’s basic algorithms invented before the 1980s No labor, no intelligence Models perform in school laboratory Updates not in real time Inadequate integration with other technologies Only highly-skilled experts can work with AI Scarcity of data scientists Training in days or even months Scarce & costly computing power Pervasive AI for all scenarios. Respects and protects user privacy Data and energy - efficient, secure, and explainable algorithms Automated / semi-automated data labeling Industrial-grade AI, perform excellently in execution Real time, closed-loop system AI as a basic skill, supported by one-stop platforms Training in minutes or even seconds Abundant & affordable computing power AS-IS TO-BE Synergy between AI and cloud, IoT, edge computing, blockchain, big data, databases, etc. Data scientists + Subject matter experts + Data science engineers Invest in AI research Develop fundamental capabilities for data & power-efficiency, secure & trusted Common capabilities for computer vision, natural language processing, decision / inference, etc. Build a full-stack AI portfolio Deliver abundant and affordable computing power Provide an efficient and easy-to-use AI platform for all scenarios between cloud, edge and device Develop an open ecosystem and talent Collaborate widely with global academic institutions, industries, and partners Investing in developing future talents for AI Strengthen existing portfolio Bring an AI mindset and techniques into existing products and solutions to create greater value and enhance competitive strengths Drive operational efficiency Apply AI to massive volumes of routine business activities for better efficiency and quality Huawei AI Strategy
3. Huawei Confidential 3 Huawei Full Stack & All Scenarios AI Portfolio CANN (Compute Architecture for Neural Networks) Ascend All Scenarios Ascend- Max Ascend- Mini Ascend- Tiny Ascend- Lite Ascend- Nano Full Stack AI Applications HiAI Service Consumer Device Public Cloud Private Cloud Industrial IoT Device Edge Computing Application Enablement Framework Chip Enablement General APIs1 Advanced APIs Pre-integrated Solutions HiAI Engine ModelArts MindSpore TensorFlow PyTorch PaddlePaddle … IP & Chip Peripherals/IO DDR/HBM ARM Cube LSU Cache/Buffer Vector Scalar N Da Vinci Core SCALABLE COMPUTE： Current control in picoseconds Hardware-assisted task scheduler Cube：4096(16^3) FP16 MACs + 8192 INT8 MACs Vector：128 (8*16) FP16 vector Scalable cube: 16x16xN， N=16/8/4/2/1 Multiple precision: int8/int32/FP16/FP32 Multiple Compute units: Tensor/Vector/Scalar SCALABLE MEMORY： Dedicated & distributed, tiling- friendly, explicit memory design 4 TByte/s L2 buffer 1.2 TByte/s HBM SCALABLE ON-CHIP INTERCONNECTION： Ultra-high bandwidth mesh network on chip
4. Huawei Confidential 4 SoftCOM AI 4 Key Areas for Service Excellence and OPEX reduction & Autonomous Network Level Definition Execution (Hands) Awareness (Eyes) Decision (Minds) Service Experience System Complexity Not applicable Sub-task Mode-specific Unit level Mode-specific Domain level Mode-specific Service level Mode-specific All modes L5: Full Autonomous Network L4: Highly Autonomous Network L3: Conditional Autonomous Network L2: Partial Autonomous Network L1: Assisted Operation & Maintenance Level Definition L0: Manual Operation & Maintenance OPEX 5.0% Maintenance Costs Employee Others Depreciation Networkt interconnect and circuit leasing Sales cost Product cost 9.5% 13.6% 12.4% 21.7% 13.4% 9.9% 14.8% Power and water heating O&M efficiency AUTIN, AI-driven operations Efficiency improvement and OPEX Savings Performance & Resource Improvement with AI Customer Experience AI-driven UX Energy Efficiency Zero bit, Zero watt
5. Huawei Confidential 5 5G Products and Optimization Energy Efficiency in DC 5G: 14nm -> 7nm AI-enabled, multi-purpose Chipset Brings Competitiveness Powerful Antenna Benefits enabled next generation services 8 Global R&D Centers 1st SG178 Antenna Test Lab No.1 Market Share World’s 1st 5G CPE World’s 1st 5G Chipset Balong 5G01 5G Smartphone 2019 H1 Kirin 980 5G Ready 5G Device Speeds Up 5G Launch * 5G Smartphone will be released in Mid-2019 From Macro to Small Cell, the Most Powerful Chip M-MIMO needs stronger antenna technology, ASIC and algorithm Terminal is very important for 5G initial commercial CLOUD INTELLIGENT UNIFIED ARCHITECTURE & PROTOCOL Network Products BBU CPE Core Transport C-Band mmWave 15% energy efficiency -8 MILLION RMB/YEAR reduced electricity bills A DC in Langfang, Hebei AI Intelligent Control Body Auto-tuning control based on AI reinforcement learning Evaluation neural networks Control neural networks PUE forecast 5 categories, 60+ environmental parameters Millions of parameter combinations Precise control Online training, closed-loop control Electricity bill 70% CAPEX, 20% Manual maintenance, 5% Rent 5% Cooling 64% Power supply and distribution 31% Others Natural cooling Natural cooling Water cooling Three years of electric bills equal the cost of deploying a new DC Cooling systems have the highest proportion of energy consumption among non-primary equipment Control value High energy consumption for large DCs, cooling system power consumption accounted for more than 60%
6. Huawei Confidential 6 5G Atlas Intelligent Computing Platform Portfolio Atlas 200 AI Accelerator Module GPU FPGA NPU Atlas G5500 Heterogeneous Server Atlas G2500 Intelligent Video Analytics Server Atlas 800 AI Appliance Atlas 300 AI Accelerator Card Atlas 500 AI Edge Station 30  200 Number of faces in a single frame Cameras Drones Robots Atlas 200 AI Accelerator Module
7. Huawei Confidential 7 Practice cases Audi and Huawei Co-Operate in Future Driving with Huawei AI Chips Huawei TaiShan: Best Performance ARM-based Server in the Industry • Huawei Ascend 910 Chipset • Huawei Mini Atlas 600 Evolving towards self-driving, zero-fault driving experience Future 2018 2017 First Time Comprehensive V2I demo at WIoT Expo 2017 (China) V2X at WIoT Expo 2018 V2I Scenarios(intergrated with ADAS) V2V Scenarios (automatic notifications/warnings) Kunpeng 920 • Huawei Research and Development • ARM-based CPU Industry Best Performance Benchmark Memory bandwidth: Total I/O bandwidth: Network bandwidth: High Performance High Throughput High Integration High Efficiency 1 Chip = 4 Chip 46% 66% 4x 930+, 25% SPECint 30% Nodes: 50% Power consumption: 33% Racks: 35%
8. Huawei Confidential 8 Best network for business success in Moscow, Keep Leading in the Future! Thank You!