Novel architectures are the next step in silicon

Adapting to the demands of modern computing

Novel architectures are the next step in silicon

Pushing the boundaries of what's possible, they have the potential to change the way we use technology in everything from business to our daily lives. But the reality of implementing this revolutionary technology isn't without its challenges.

Traditional chips are struggling to keep up with the enormous computational resources required of modern computing, leading the trajectory of silicon design down a new path as the long-held theories of Moore's Law and Dennard Scaling approach their limits. Instead, the future of silicon design lies in novel silicon architectures.

What are novel silicon architectures?

The ever-increasing call for more intense computation has highlighted the limitations of traditional chips. For decades transistors have continued to shrink, but this trend falls short when confronted with the complex AI algorithms and workloads modern computing demands.

Despite ongoing efforts to boost chip density, progress has slowed, as has the reduction in power consumption per unit, resulting in larger chips that run the risk of overheating.

Novel silicon architectures can take on these challenges. Facilitating high speed, improved performance and energy efficiency, these specialised chips can keep up with the demands of modern computing with ease, opening the doors to a myriad of applications.

To allow for this impressive performance, novel silicon architectures don't solely rely on ever-decreasing transistor size. Instead, they follow three key approaches.

Multi-layered chiplet-based designs are developed to address a specific task. They are made up of multiple chiplets been specially designed to fit the required criteria, and then stacked into a 3D design.

One of the benefits of this approach is that chiplets with different-sized transistors can be mixed and matched to achieve the best energy efficiency and cost-effectiveness for the task at hand. These designs can also maintain high density while using well-established technology without relentlessly pushing transistor size down to its extreme, allowing for more reliability and reduced production costs.

Parallel processing, on the other hand, excels at performing multiple different tasks simultaneously to achieve fast and refined results. With the ability to increase computation speed while maintaining or even decreasing clock speed, parallel processing can handle large amounts of computational data while maintaining low latency and incredible energy efficiency.

Analogue compute-in-memory (CIM) also excels at maintaining ultra-low power and latency, achieving this by combining analogue circuits with compute-in-memory. This makes it well-suited to low-power edge devices as it isn't slowed down by the latency of fetching data from an external memory, allowing it to efficiently process computationally intensive AI workloads.

Whether used independently or in combination, each approach has unique strengths that support both AI and other challenging requirements in a more adaptable model.

Applications of novel silicon architectures

Thanks to their versatile nature, high computation speed and low power consumption, these approaches can be employed for a vast range of applications within modern computing.

Neuromorphic computing is an important area where novel silicon architectures are essential. By making use of parallel processing, neuromorphic computing can imitate the efficiency and elegance of the human brain in processing immense amounts of data, all while maintaining low latency and power consumption.

In turn, neuromorphic computing can be utilised for applications like myoelectric prosthetics, providing precise and immediate processing without increasing power consumption to provide the best user experience. We explore these use cases further in our recent whitepaper, ‘Future of computing: the new commercial horizon.'

The UAV (uncrewed aerial vehicles) market is another area that could use a novel silicon architecture to its advantage. With the low-power capabilities we've discussed, drones would be better equipped to take on extended missions for tasks like agricultural monitoring or disaster assessment, all while running complex AI algorithms locally and feeding back to a control station.

These are just a few examples to illustrate the many opportunities novel silicon architectures present. In the years to come, it's clear to me that they'll continue to enable silicon technology to evolve, keeping up with the dramatic advancements of AI and other emerging computing trends.

null

Dr Aidong Xu is Head of Semiconductor Capability at Cambridge Consultants. He has over 30 years' experience across diverse industries, including with leading semiconductor companies. He has managed large international engineering teams and brought products into the global market that have achieved rapid and sustained business growth. Aidong holds a PhD. in power electronics and power semiconductor.