深度學習芯片組的全球市場：用於AI學習/推理的CPU / GPU / FPGA / ASIC / SoC加速器

人工智能（AI）加速的需求已得到廣泛認可，這使得AI加速芯片組成為企業（數據中心）和邊緣設備製造商的標準功能要求。結果，在過去兩年中，AI芯片組的出貨量和收入已顯示出顯著增長。全球深度學習芯片組收入預計將從2019年的114億美元增長到2025年的712億美元。

該報告調查了深度學習芯片組的市場，市場定義和概述，分析了影響市場增長的各種因素，技術趨勢，加速需求，芯片組要求。 ，收入規模的轉變和預測，按芯片組類型，學習/推理類別，計算能力，最終用戶，競爭環境，主要公司的概況等各種類別細分。

執行摘要

市場分析

• 在市場上使用AI
  • 企業中的AI加速
  • 邊緣的AI加速
• 市場細分
  • 按體系結構（芯片組類型）
  • 學習與推理
  • 通過計算能力
  • 按功耗
  • 按最終用戶：Enterprise Edge
• 市場增長因素
• 市場障礙/問題
• 應用程序/用法示例
• 地區差異
• 啟動業務計劃

技術分析

• 神經網絡演進與硬件加速需求
• AI推理工作量
• 神經網絡解釋和芯片組要求
• 用於深度學習的芯片組架構
• 新的AI加速架構
• 芯片組技術參數
• 基準化
• 用於AI芯片組的數據格式
• 深度學習開發框架

主要公司

• Amazon
• AMD
• ARM
• Cerebras Systems
• CEVA
• Esperanto Technologies
• Facebook
• Google
• Graphcore
• Groq
• Gyrfalcon Technologies
• Habana Labs (Intel)
• Huawei
• Intel
• Kalray
• MediaTek
• Movidius (Intel)
• Mobileye (Intel)
• NVIDIA
• Qualcomm
• SambaNova
• Thinci (now Blaize)
• Xilinx

市場預測

• 預測方法/前提條件
• 總市場
• 按芯片組類型劃分的收入
• 利潤：通過學習和推理
• 按計算能力計算的收入
• 功耗消耗
• 平均售價：按芯片組類型
• 最終用戶的收入
• CPU
• GPU
• ASIC
• FPGA
• SoC加速器
• 概述

The need for artificial intelligence (AI) acceleration is widely recognized as of 2020. AI acceleration chipsets have become a standard feature requirement for device manufacturers within the enterprise (data center) and edge markets. As a result, the volume and revenue of AI chipsets have increased drastically in the last two years. NVIDIA's latest A100 offers petaOPS of compute performance under certain compute conditions, marking a tremendous jump from the petaOPS server DGX-1 introduced just two short years ago.

Deep learning (DL) is slowly moving past its hype cycle as proof-of-concept (PoC) AI applications developed in the past two years go into production. AI chipset customers have become more sophisticated in terms of chipset needs for AI application acceleration and are asking for specific benchmarks when talking to vendors. Customers' needs for chipsets are coming to the forefront, forcing chipset companies to rethink the applicability of their technology. All prominent chip companies, such as Intel, NVIDIA, and Qualcomm, have invested heavily in AI. Cloud companies have started rolling out graphics processing units (GPUs), field-programmable gate arrays (FPGAs), and application-specific integrated circuits (ASICs), giving developers a choice for AI acceleration. Omdia forecasts that global revenue for DL chipsets will increase from $11.4bn in 2019 to $71.2bn by 2025.

This Omdia Market Report assesses the industry dynamics, technology issues, and market opportunity surrounding DL chipsets, including CPUs, GPUs, FPGAs, ASICs, and SoC accelerators. As an update to Omdia's 2019 ‘Deep Learning Chipsets ’ report, it captures the state of this fast-moving chipset market. Global market forecasts, segmented by chipset type, compute capacity, power consumption, market sector, and training versus inference, extend through 2025. Omdia also provides profiles of 23 key industry players.

Key Questions Addressed:

• What chipset types are being used for deep learning (DL) today, and how will they change through 2025 and beyond?

• What are the power consumption and compute capacity profiles of chipsets used for DL applications?

• What is the market opportunity for DL chipsets in enterprise environments versus edge devices?

• Which market sectors and industries will drive demand for DL chipsets?

• What is the state of technology development for DL chipsets, and which companies are driving innovation?

• What are some of the emerging architectures for DL chipsets?

• What are the key performance matrices for DL chipsets?

• What are some of the use cases for DL chipsets in different application markets?

• What has changed in the DL chipset market in the last two years?

• How are startups faring in the DL chipset market?

Who Needs This Report?

• Semiconductor and component manufacturers
• OEM companies building devices using AI chipsets
• Cloud companies using AI chipsets
• Service providers and systems integrators
• End-user organizations deploying deep learning systems
• Industry associations
• Government agencies
• Investor community

Table of Contents

Executive summary

• Introduction
• 2020 report update
• Key findings
• Market forecasts

Market issues

• Use of AI in the market
  • AI acceleration within enterprises
  • AI acceleration at the edge

• Market segmentation
  • Segmentation by architecture (chipset type)
  • Segmentation based on training vs. inference
  • Segmentation based on compute capacity
  • Segmentation based on power consumption
  • Segmentation based on market sector: Enterprise and edge market

• Market drivers
  • Popularity of AI and increasing complexity
  • Multiple AI pipelines
  • Complexity of training
  • Growth in enterprise applications
  • Desire to minimize production costs
  • Latency and throughput requirements for inference
  • Computer vision
  • Speech applications for embedded devices

• Market barriers and challenges
  • Capital needs for chip development
  • Availability of expertise
  • Long development cycle and rapidly changing market

• Applications and use cases
  • Enterprise applications and use cases
  • Edge applications and use cases
  • Other

• Regional differences

• Startup activity in deep learning chipsets
  • Many acquisitions
  • Casualties

Technology issues

• Evolution of neural networks since 2012 and the need for hardware acceleration
  • Computation needs per forward pass (inference)
  • Compute needs for training
  • A neural network zoo

• AI inference workloads
  • Recommendation engine
  • Image and video
  • Audio and speech
  • Text/natural language processing
  • Search

• Translating neural network needs to chipset requirements
  • Processing elements and arithmetic logic units
  • Memory
  • On-chip connectivity
  • Chip-to-chip connectivity

• Chipset architectures for deep learning
  • Central processing units
  • Graphics processing units
  • Field-programmable gate arrays
  • Application-specific integrated circuits
  • System-on-chip accelerators

• Emerging AI acceleration architectures
  • Optical computing
  • Analog computing
  • Processing in memory
  • Neuromorphic

• Technology parameters for chipsets
• Benchmark
• Data formats used in AI chipsets
• Deep learning development frameworks

Key industry players

• Amazon
• AMD
• ARM
• Cerebras Systems
• CEVA
• Esperanto Technologies
• Facebook
• Google
• Graphcore
• Groq
• Gyrfalcon Technologies
• Habana Labs (acquired by Intel)
• Huawei
• Intel
• Kalray
• MediaTek
• Movidius (Intel)
• Mobileye (Intel)
• NVIDIA
• Qualcomm
• SambaNova
• Thinci (now Blaize)
• Xilinx
• Deep learning chipset and IP companies

Market forecasts

• Forecast methodology and assumptions
  • Omdia coverage of AI chipsets

• Overall market
• Revenue by chipset type
• Revenue by training vs. inference
• Revenue by compute capacity
• Revenue by power consumption
• Average selling price by chipset type
• Revenue by market sector
• Central processing units
• Graphics processing units
• Application-specific integrated circuits
• Field-programmable gate arrays
• System-on-chip accelerators
• Conclusions

Tables

• Deep learning chipset revenue by chipset type, world markets: 2019-25
• Deep learning chipset revenue, enterprise vs. edge, world markets: 2019-25
• Deep learning chipset revenue growth rates, world markets: 2020-25
• Deep learning chipset revenue by power consumption, world markets: 2019-25
• Deep learning chipset revenue by compute capacity, world markets: 2019-25
• Deep learning chipset revenue, inference vs. training, world markets: 2019-25
• Deep learning edge chipset revenue by chipset type, world markets: 2019-25
• Deep learning edge chipset shipments by chipset type, world markets: 2019-25
• Deep learning edge chipset revenue growth rates, world markets: 2020-25
• Deep learning edge chipset revenue by power consumption, world markets: 2019-25
• Deep learning edge chipset revenue by compute capacity, world markets: 2019-25
• Deep learning edge chipset revenue for inference vs. training, world markets: 2019-25
• Deep learning enterprise chipset revenue by chipset type, world markets: 2019-25
• Deep learning enterprise chipset shipments by chipset type, world markets: 2019-25
• Deep learning enterprise chipset revenue growth rates, world markets: 2020-25
• Deep learning enterprise chipset revenue by power consumption, world markets: 2019-25
• Deep learning enterprise chipset revenue by compute capacity, world markets: 2019-25
• Deep learning enterprise chipset revenue for inference vs. training, world markets: 2019-25
• Deep learning edge chipset ASPs by chipset type, mobile, HMDs, drones, and machine vision (non-PC), world markets: 2019-25
• Deep learning edge chipset ASPs by chipset type, edge servers, world markets: 2019-25
• Deep learning edge chipset ASPs by chipset type, PCs/tablets, world markets: 2019-25
• Deep learning edge chipset ASPs by chipset type, cameras, world markets: 2019-25
• Deep learning edge chipset ASPs by chipset type, smart speakers, world markets: 2019-25
• Deep learning edge chipset ASPs by chipset type, automotive, world markets: 2019-25
• Deep learning edge chipset ASPs by chipset type, robots, world markets: 2019-25
• Deep learning enterprise training chipset ASPs by chipset type, world markets: 2019-25
• Deep learning enterprise inference chipset ASPs by chipset type, world markets: 2019-25
• Deep learning CPU chipset revenue by market sector, world markets: 2019-25
• Deep learning CPU enterprise chipset revenue, training vs. inference, world markets: 2019-25
• Deep learning CPU edge chipset revenue, training vs. inference, world markets: 2019-25
• Deep learning CPU enterprise chipset revenue by power consumption, world markets: 2019-25
• Deep learning CPU enterprise chipset revenue by compute capacity, world markets: 2019-25
• Deep learning CPU edge chipset revenue by power consumption, world markets: 2019-25
• Deep learning CPU edge chipset revenue by compute capacity, world markets: 2019-25
• Deep learning GPU chipset revenue by market sector, world markets: 2019-25
• Deep learning GPU enterprise chipset revenue, training vs. inference, world markets: 2019-25
• Deep learning GPU edge chipset revenue, training vs. inference, world markets: 2019-25
• Deep learning GPU enterprise chipset revenue by power consumption, world markets: 2019-25
• Deep learning GPU enterprise chipset revenue by compute capacity, world markets: 2019-25
• Deep learning GPU edge chipset revenue by power consumption, world markets: 2019-25
• Deep learning GPU edge chipset revenue by compute capacity, world markets: 2019-25
• Deep learning FPGA chipset revenue by market sector, world markets: 2019-25
• Deep learning FPGA enterprise chipset revenue, training vs. inference, world markets: 2019-25
• Deep learning FPGA edge chipset revenue, training vs. inference, world markets: 2019-25
• Deep learning FPGA enterprise chipset revenue by power consumption, world markets: 2019-25
• Deep learning FPGA enterprise chipset revenue by compute capacity, world markets: 2019-25
• Deep learning FPGA edge chipset revenue by power consumption, world markets: 2019-25
• Deep learning FPGA edge chipset revenue by compute capacity, world markets: 2019-25
• Deep learning ASIC chipset revenue by market sector, world markets: 2019-25
• Deep learning ASIC enterprise chipset revenue, training vs. inference, world markets: 2019-25
• Deep learning ASIC edge chipset revenue, training vs. inference, world markets: 2019-25
• Deep learning ASIC enterprise chipset revenue by power consumption, world markets: 2019-25
• Deep learning ASIC enterprise chipset revenue by compute capacity, world markets: 2019-25
• Deep learning ASIC edge chipset revenue by power consumption, world markets: 2019-25
• Deep learning ASIC edge chipset revenue by compute capacity, world markets: 2019-25
• Deep learning SoC accelerator chipset revenue by market sector, world markets: 2019-25
• Deep learning SoC accelerator enterprise chipset revenue, training vs. inference, world markets: 2019-25
• Deep learning SoC accelerator edge chipset revenue, training vs. inference, world markets: 2019-25
• Deep learning SoC accelerator enterprise chipset revenue by power consumption, world markets: 2019-25
• Deep learning SoC accelerator enterprise chipset revenue by compute capacity, world markets: 2019-25
• Deep learning SoC accelerator edge chipset revenue by power consumption, world markets: 2019-25
• Deep learning SoC accelerator edge chipset revenue by compute capacity, world markets: 2019-25
• Types of devices with enterprises using AI accelerators
• Edge devices shipping in high volume and chipset requirements
• Key players in different deep learning chipsets
• Key CPU products and vendors
• Key players in GPU
• Key players in FPGA
• Comparison of deep learning chipset parameters
• Selected benchmarks for AI chipsets
• Data formats used in AI chipsets
• Popular deep learning frameworks
• Deep learning chipset companies
• IP companies

Figures

• Deep learning chipset revenue, world markets: 2019-25
• Estimated AI workloads on enterprise GPUs and CPUs
• Deep learning chipset revenue, world markets: 2019-25
• Deep learning chipset year-on-year revenue growth rates, world markets: 2020-25
• Deep learning chipset revenue by chipset type, world markets: 2019-25
• Deep learning chipset revenue, inference vs. training, world markets: 2019-25
• Deep learning chipset revenue by compute capacity, world markets: 2019-25
• Deep learning chipset revenue by power consumption, world markets: 2019-25
• Deep learning chipset revenue by market sector, world markets: 2019-25
• Deep learning CPU chipset revenue, world markets: 2019-25
• Deep learning GPU chipset revenue, world markets: 2019-25
• Deep learning ASIC chipset revenue, world markets: 2019-25
• Deep learning FPGA chipset revenue, world markets: 2019-25
• Deep learning SoC accelerator chipset revenue, world markets: 2019-25