人工智能芯片发展前景研究III_关键机会与趋势（英文版）_22页_1mb

AI-Optimized Chipsets: Summary

Core Content

The AI-optimized chipset market is segmented into training and inference markets, each with distinct characteristics and players. The training market is dominated by a few key firms, while the inference market is more diverse and accessible to a wide range of companies, including startups and tech giants.

Market Overview

Key Players and Strategies

Training Chipset Market

• Nvidia is the dominant player, with a strong presence in the data center (DC) and GPU technology stack. It offers both training and inference solutions.
• Intel, AMD, and Google also participate in the training market, with Google developing its own TPU (Tensor Processing Unit) for AI tasks.

Inference Chipset Market

• The inference market is more diverse, with over 40 companies involved.
• CPU is the de facto choice for inference in DCs.
• FPGAs are gaining traction, with Microsoft, Amazon, and Alibaba using them for inference tasks.
• ASICs are being developed by startups such as Graphcore, Wave Computing, Cerebras, Groq, and Tenstorrent to target specific AI applications.

Edge vs. Cloud

• Edge computing is expected to represent more than 75% of the total AI chipset market opportunity.
• Training is primarily conducted in DCs, while inference is moving towards the edge, including applications in automotive, drones, smart cameras, and mobile devices.
• NVIDIA, Xilinx, and Qualcomm are key players in both edge and cloud inference markets.

Tech Giants' Involvement

• Super 7 Hyperscalers (Alibaba, Amazon, Baidu, Facebook, Google, Microsoft, Tencent) are increasingly building proprietary AI-optimized chipsets due to their hyperscale data center expertise.
• These firms are not only developing their own chipsets but also acquiring startups to enhance their capabilities:
  • Intel acquired Altera (FPGA maker) and C-SKY (CPU designer)
  • Xilinx acquired DeePhi Tech
  • Qualcomm acquired NXP (for automotive applications)

Startups and Innovations

• Graphcore (UK) develops IPUs (Intelligence Processing Units) with a graph-based architecture and 100x more memory bandwidth than other solutions.
• Wave Computing (US) focuses on DPU (Dataflow Processing Unit) with low power and cost and no host CPU dependency.
• Cerebras (US) is developing a specialized training chip for sparse matrix operations.
• Groq (US) claims to deliver 400 teraflops performance, more than twice Google's TPU.
• Tenstorrent (Canada) is working on adaptive computation and scalable AI chips.
• Horizon Robotics (China) and Mythic (US) are developing ASICs for edge computing with low power and high performance.
• Cambricon (China) offers AI chips for both cloud and edge, with a software platform for developers.
• Kneron (China, US, Taiwan) provides RANN (Reconfigurable Artificial Neural Network) technology for edge AI solutions.

Challenges and Risks

• Execution Risks: It may take years for startups to bring their chipsets to market, and many are still in development.
• Narrow Focus: Startups that focus too narrowly on specific applications may fail if those applications do not gain mainstream traction.
• Market Saturation: The AI-optimized chipset market is super-saturated, and it is unclear where the exits will be for most players.

Network Effects and Switching Costs

• GPUs have a strong ecosystem and developer support, making them difficult to replace.
• Switching costs for end users are high, especially for those already invested in GPU-based systems.
• Google continues to offer GPU access via its cloud services, despite its own TPU development.

Conclusion

• The AI-optimized chipset market is evolving rapidly, with a mix of established players and disruptive startups.
• Edge computing is becoming a major driver for AI chip development.
• Nvidia remains a leader in both training and inference markets, while Intel, Xilinx, and Google are also making significant strides.
• The market is expected to expand significantly in 2019 and 2020, with winners emerging from the competition.