AI Series II: ASIC, Customized Chips Powering Intelligent Takeoff

In our previous article “AI Series I: Who Is Laying the Foundation For AI?”, we explored the importance of artificial intelligence infrastructure and how companies like Lam Research, Applied Materials, and Taiwan Semiconductor Manufacturing Company have laid a solid foundation for the development of AI. Today, we will focus on the ASIC (Application-Specific Integrated Circuit) industry, the customized chips that are becoming the backbone of AI’s next leap forward.

ASIC Chips: Customized "Accelerators" for AI

Imagine a world where your smartphone, smart speaker, and even your car's computer system are all powered by chips designed specifically for the tasks they perform. This is the world that ASICs are creating. Application-Specific Integrated Circuits, as the name suggests, are integrated circuits designed for specific applications. Unlike GPUs, which are like Swiss Army knives capable of many tasks but not always the best at any one, ASICs are like specialized tools crafted for a single purpose. They are the reason your voice commands are understood instantly, and why complex AI models can run efficiently on your devices.

Let's take a step back and think about how we use AI today. From the large language models that power chatbots to the computer vision systems in self-driving cars, AI is everywhere. But these applications have very different needs. A chip that works well for one might be a poor fit for another. This is where ASICs shine. They are designed from the ground up for specific tasks, making them more powerful and energy-efficient than GPUs, which are jacks-of-all-trades but masters of none.

For example, when it comes to training large language models, GPUs might struggle to keep up. But an ASIC designed for this task can handle the massive amounts of data and complex calculations with ease. This means faster training times and lower energy costs, a win-win for both developers and the environment.

Applications of ASIC Chips

ASIC chips are not just theoretical concepts; they are already making a significant impact in various AI applications.

In the field of deep learning, ASIC chips can significantly improve the efficiency of training and inference. As early as 2015, Google developed the first-generation TPU (Tensor Processing Unit) processor, which powered AlphaGo's victory over the top Go player Lee Sedol. TPU is an ASIC chip specially designed for machine learning. By optimizing the hardware for machine learning models, the computing performance of TPU is greatly enhanced. Therefore, TPU can provide higher performance and lower power consumption than traditional GPUs when processing deep learning workloads. This enables Google to run its machine learning models in data canters more efficiently, providing users with faster and more accurate services.

In the field of natural language processing, Amazon Web Services has tailored the Trainium series chips by optimizing their design and scaling their application. They have developed specialized circuit modules for operations like convolution and matrix multiplication. During inference tasks, these chips minimize redundant calculations and unnecessary circuit switching, resulting in lower power consumption and reduced costs.

In the field of computer vision, ASIC chips can accelerate the processing of images and videos. For example, some ASIC chips designed specifically for computer vision can provide higher frame rates and lower power consumption when processing video streams in real-time. This enables applications such as drones and self-driving cars to operate more efficiently.

‍

Top Players in the ASIC Industry

As the development of artificial intelligence enters the outbreak period of application level, the demand for computing power is gradually shifting from training to inference. Since the requirements for chips in inference are lower, Nvidia's GPU no longer has absolute dominance, and the low-cost solutions of ASIC chips for lower inference computing power will be more suitable for AI companies. Therefore, we expect that ASIC chips will enter a period of rapid growth from 2025 to 2027. ASIC manufacturers led by Broadcom are expected to achieve a 70% compound annual growth rate, higher than the growth rate of traditional chips.

• Broadcom: Broadcom's AI business is mainly divided into ASIC and switch categories, with ASIC accounting for about two-thirds of the company's business. Broadcom announced at the end of 2024 that its AI business revenue will reach 60-90 billion US dollars in 2027. In addition, Broadcom has established cooperation with market leaders such as Google, Meta, OpenAI, and ByteDance to develop the next generation of artificial intelligence chips.
• Marvell: As a challenger in the ASIC chip field, Marvell's current ASIC revenue mainly comes from its cooperation with Amazon's Trainium project, and the company's cooperation with Amazon's Inferential project will also begin mass production in 2025.
• Nvidia: Although Nvidia's investment in ASIC chips is later than that of main market participants such as Broadcom and Marvell, Nvidia's advantage in full-link design capabilities will also help it quickly occupy a favorable position. Nvidia's GPU has outstanding performance in AI training tasks, and it may gradually enter the ASIC chip field in the future, providing customized solutions for specific applications.

‍

The Road Ahead: Challenges and Opportunities

As we look to the future, the potential of ASICs is immense. They are the key to unlocking the next level of AI performance, from faster and more efficient AI models to new applications we can only begin to imagine.

Future Trends

• Performance Improvement: With the continuous advancement of manufacturing process technology, ASIC chip products have evolved to 3nm and 5nm, significantly enhancing computing power and energy efficiency.
• Growing Customization Demand: Different artificial intelligence applications have different requirements for chip performance and functions. ASIC chips can be customized for specific applications to meet the needs of different users. Currently, the business models and application scenarios of Cloud Service Providers like Amazon are mostly carried through their own clouds, and ASIC chips can be developed targetedly according to the needs of different business scenarios and commercial models.
• System Integration: Referring to the development history of Nvidia, future ASIC chips will not only compete purely on computing power or cost but also require chip manufacturers to accumulate in chip design, networks, and interfaces, focusing more on system advantages. This advantage makes chips more competitive in terms of performance, power consumption, and cost.

However, although ASIC chips can help AI companies reduce chip costs, the research and development costs of ASIC chips are not low at present. This is because the design and manufacturing stages of ASIC chips require a large amount of research and development investment, and the development cycle is very long. For example, Google's TPU chip started research and development in 2015, and it was not until 2022 that chips for machine learning appeared. In addition, compared with Nvidia's GPU, both Broadcom and Marvell lack a rich and easy-to-use software ecosystem like CUDA, making it difficult to modify ASIC chips once they are produced.

Yet, the rewards are worth the challenges. With the right investments and innovations, ASICs could become the standard for AI computing, driving a new wave of technological advancement.

Conclusion

In conclusion, ASICs are not just another type of chip; they are critical components of the future development of AI. They represent a shift from one-size-fits-all solutions to a world where every application has a chip designed specifically for it. This customization is what will allow AI to reach its full potential, from the smartphones in our pockets to the self-driving cars on our roads.

As we continue to explore the possibilities of AI, let's keep an eye on the ASIC industry. It is here that the real magic happens, where ideas become reality, and where the future of AI is being built, one chip at a time. Stay tuned for more insights in our ongoing AI series, where we'll continue to uncover the innovations shaping the world of tomorrow.

‍