Software is taking over the world, and now artificial intelligence (AI)a branch of computer science that lets machines imitate human reasoning- is taking over software. In the last decade, machine learning models (systems that learn patterns from data to make decisions) have become much more complex. As a result, running models for tasks such as image and voice recognition or translation now requires much more computing power. Some models require more than a petaflop per second to run, meaning doing a thousand trillion calculations every second for a full day.
So, how are the chips powering these models keeping up?
As AI applications require more computing power, a new trend is emerging: AI is now used to design the very chips that enable them. This development marks a significant shift from the industry’s origins and highlights the evolving impact of AI on hardware design. To clarify how this shift is unfolding, I will explain how AI has moved our focus from software to hardware and why this evolution matters for electronic design automation (EDA) technologies.
AI Brings New Opportunities—and Challenges—to Chip Designers
Software has long set tech leaders apart, but with AI’s rise, the focus now shifts decisively to hardware as the backbone of technological innovation. AI and its learning methods—machine learning and deep learning are driving this shift by enabling machines to perform complex, human-like tasks. These advances demand new hardware solutions that can keep pace with AI’s needs, moving the competition and innovation frontier to chip design.
As intelligent systems from virtual assistants to self-driving cars proliferate, their demand for powerful, real-time data processing is driving rapid growth in AI-related semiconductor markets. Significant market opportunities and value are now concentrated in the hardware layer, attracting new entrants into advanced chip development and reinforcing the view that the future of AI hinges on breakthroughs in chip design.
The increasing demand for powerful chips is a part of a longer history. AI has been around since the 1950s. The math from the early days still matters, but back then, we couldn’t use AI in everyday life. In the 1980s, expert systems emerged, performing tasks such as matching symptoms on healthcare websites. Deep learning arrived in 2016, bringing big changes like image recognition and making hardware performance more important. Now, AI is being used in more than just big systems like cars or scientific models. It’s moving from data centers and the cloud to the edge, where trained neural networks make decisions about new data based on what they’ve already learned.
Following this trend, devices like smartphones, AR and VR headsets, robots, and smart speakers now use AI at the edge, meaning processing occurs on the device itself. By 2025, experts expect seventy percent of AI software to run this way, with hundreds of millions of edge AI devices already in use. We’re seeing a huge increase in real-time data processing, often requiring twenty to thirty models and only microseconds of delay. For example, self-driving cars or drones need to respond in just twenty millionths of a second for safety. Voice and video assistants need even faster responses, dash under ten milliseconds for recognizing keywords and under one millisecond for hand gestures.
Take Google’s LTLSTM1 voice recognition model, for example. It uses natural language, has 56 layers and 34 million weights, and performs about 19 billion operations per guess. To work well, it must understand a question and answer in less than 7 milliseconds. To achieve this, Google created its own chip, the Tensor Processing Unit (TPU). Now, in its third generation, the TPU demonstrates how new hardware needs are driving new hardware designs, helping speed up neural network tasks across many Google services. Application-specific optimizations cannot yet compete with human capabilities. But there’s more on the horizon. In the research phase, instances are:
- Neuromorphic computing, a type of computer architecture designed to work like the human brain, provides an intrinsic understanding of a problem within a model and examines thousands of characteristics to deliver ultimate parallelism (the ability to process many tasks at once).
- Another area of research is high-dimensional computing, where patterns are learned using single-shot learning methods (which enable systems to recognize new patterns or objects based on just one or a few examples).
Though promising, these research areas are still far from efficiently handling such computing tasks on today’s chips. Still, semiconductor advances are underway and will eventually change this situation.
Spearheading the Era of Autonomous Chip Design
Echoing the main argument, AI is now both the consumer and the creator of next-generation chips. With tools like DSO.ai, AI uses reinforcement learning to autonomously navigate complex chip design decisions, dramatically improving performance and efficiency. As industry leaders adopt these tools, autonomous design is solidifying AI’s transformative impact on the entire hardware ecosystem.
Another part of this shift is the need for faster, more flexible chip design and manufacturing. Designing a chip takes one or two years, and large-scale manufacturing takes longer. Designers must make chips adaptable for useful applications years after they are planned. The industry suggests software-defined hardware chips reprogrammable post-manufacturing to balance flexibility and performance. Tools like DSO.ai enable this much faster and more cost-effectively than humans alone.
Looking ahead, it’s possible that AI will help achieve the next 1,000-fold increase in computing power, which the industry will need as more devices and systems get smarter. This is an exciting time, with a new approach that uses software to guide the entire hardware design process, optimizing how systems work, how they’re built, and how they’re placed on chips. And all of this happens much faster and with less engineering effort than before.
In summary, we’ve learned that the era of autonomous chip design spans everything from circuit simulation, layout, and verification to digital simulation, synthesis, IP reuse, and customer hardware solutions. AI-driven design tools are pushing the limits of what chips can do, which is essential for meeting the needs of AI applications. It’s a lucky cycle that makes this an especially exciting time to work in electronics.
Source: AI Chip Design Enables Breakthroughs for Chip Makers
