Summary

One of the key factors behind the rise of artificial intelligence (AI) is the upgrade in cloud computing power. This is largely driven by the enhancement and upgrade in cloud AI chipsets. Cloud AI chipsets are computational chipsets focusing on AI workload that is typical deployed in the cloud, or data center, environment. This chipset can be designed specifically for AI inference or AI training. In some instances, the chipset can support both AI inference and AI training.

Due to the constant evolution of AI algorithms, cloud AI chipsets are designed to support wide range of AI models, from rule-based AI to deep learning models, with varying degree of resource requirements. As compared to edge AI chipsets, a cloud AI chipset generally has higher computational power, higher power consumption, larger physical footprint and is therefore relatively more expensive.

Cloud AI market is so far dominated by NVIDIA GPUs and Intel’s CPUs. In recent years, many companies have started to emerge and offer interesting take on how to address the challenge of AI workload in the cloud. On one hand, new startups like Cerebras Systems, Graphcore, Habana Labs, and Wave Computing have announced new chipsets that have higher performance or better computational flow as compared to conventional chipsets. On the other hand, captive vendors have started to build their own AI chips to power their data centers. Examples of these vendors include Amazon, Google, Huawei, Baidu and potentially Alibaba.

Overall, the market size for cloud AI chipsets is expected to be US$3.5 billion in 2018. This is expected to grow to US$19.1 billion in 2024. Right now, most of the market share is captured by non-captive vendors. As cloud service providers are going to take away majority of the AI workloads, we believe that their market share will grow from 2.3% in 2018 to 9.4% in 2024. For companies to be successful in this sector, the chipset must be highly scalable and flexible, achieve the right balance between performance and power budget, but also feature strong ecosystem support and comprehensive software stack.

ページTOPに戻る

Table of Contents

• 1. EXECUTIVE SUMMARY
• 2. DEFINITION OF ARTIFICIAL INTELLIGENCE
• 3. THE EVOLUTION OF CLOUD AI CHIPSET
  • 3.1. General-Purpose AI Hardware
  • 3.2. The Rise of the FPGA
  • 3.3. Custom Chips for Specific Workloads
  • 3.4. Heterogenous Computing
• 4. DEFINITIONS OF CLOUD AI CHIPSETS
• 5. MAJOR CLOUD AI CHIPSETS
  • 5.1. Cloud AI Chipsets for Inference
  • 5.2. Cloud AI Chipsets for Training
• 6. THE RISE OF CAPTIVE VENDORS
• 7. KEY VENDOR PROFILES
  • 7.1. AWS
  • 7.2. Baidu
  • 7.3. Bitmain
  • 7.4. Cambricon Technologies
  • 7.5. Cerebras Systems
  • 7.6. Google
  • 7.7. Graphcore
  • 7.8. Habana Labs
  • 7.9. HiSilicon
  • 7.10. Intel
  • 7.11. NVIDIA
  • 7.12. Qualcomm
  • 7.13. Xilinx
  • 7.14. Wave Computing
• 8. MARKET FORECASTS
  • 8.1. Cloud AI Training versus Inference
  • 8.2. Cloud AI Chipset Architecture
  • 8.3. The Rise of Captive Vendors and Their Impact on Vendor Share
• 9. KEY RECOMMENDATIONS AND CONCLUSIONS