ai’s next act: genius chips, programmable silicon and the future of computing
Consider this: if only a small fraction of the global population possessed the means to operate a mobile phone, would that technology have revolutionized society as profoundly as it has?
A common observation is that the innovations of tomorrow are already in existence, though their benefits aren't universally accessible. This is particularly relevant when considering the realms of artificial intelligence (AI) and machine learning (ML). Numerous sophisticated AI/ML applications are currently available, yet many demand substantial computational resources – resources typically found only with the largest corporations or at the level of national governments.
Should significant advancements in the efficiency of silicon architecture not materialize, AI’s potential will remain unequally shared, and a large portion of the world’s population will be excluded from the advantages AI could offer.
The subsequent phase in technological development hinges on finalizing the shift towards silicon architecture that is as adaptable, effective, and fundamentally programmable as the software systems we currently utilize. Failure to substantially broaden access to ML will result in a considerable loss of potential innovation, concentrating control of vital technologies within a limited number of organizations. Therefore, it’s crucial to understand what changes are necessary, the speed at which these changes are occurring, and the implications for the future of technology.
The increasing accessibility of AI: A positive development for startups and small companies
Individuals employed by large, established corporations – even those not traditionally considered technology companies – may not find the challenges with existing AI and machine learning computing resources particularly pressing.
However, for those operating with limited funding or personnel, the following forecasts signal the beginning of a transformative period where organizations of every scale and financial standing will be able to utilize equally sophisticated AI and ML-driven applications. Similar to how mobile phones broadened access to the internet, the industry is currently shifting towards making AI available to a wider audience.
Naturally, this increased accessibility requires substantial technological progress to genuinely simplify AI implementation – simply wanting to make AI more available isn't sufficient, despite the valuable efforts of companies such as Intel and Google. Several technological developments will contribute to this outcome.
From basic processor to intelligent chip
Historically, the primary measure of a processor’s quality was its raw processing speed, and their construction reflected this focus. As software became increasingly prevalent, processors needed to evolve, becoming more efficient and readily available. This led to the development of specialized processors, such as GPUs – essentially, more advanced or “smart” chips.
These graphics processing units, as a fortunate outcome, demonstrated greater effectiveness than CPUs when applied to deep learning tasks, establishing the GPU as a crucial component in contemporary artificial intelligence and machine learning. Considering this progression, the subsequent stage in development is logical: if we can design hardware specifically for graphics applications, why not also tailor it for particular deep learning, AI, and ML functions?
Several converging factors are making the coming years particularly significant for both chip manufacturing and the broader technology landscape. Firstly and secondly, we are observing a slowing down of Moore’s Law (the prediction of transistor density doubling approximately every two years) and the conclusion of Dennard scaling (which posited a similar doubling of performance-per-watt). Previously, these principles meant that each new technology generation would bring a doubling of chip density and processing power without increasing energy consumption. However, we have now reached the nanoscale, encountering the inherent limits of physics.
Thirdly, adding to these physical constraints, the computational requirements of cutting-edge AI and ML applications are exceeding previous expectations. For instance, training neural networks to achieve even a modest level of human-like image recognition is remarkably challenging and demands substantial processing capabilities. The most demanding machine learning applications include natural language processing (NLP), recommendation engines handling vast numbers of options, and extremely high-resolution computer vision utilized in fields like medicine and astronomy.
Even anticipating the need to develop and train algorithmic systems to understand human language or recognize objects in space, we underestimated the sheer amount of training – and therefore processing power – required for these models to become genuinely useful and “intelligent.”
Currently, only a select number of organizations are capable of undertaking these complex ML applications. These are typically leading businesses or research institutions with access to significant computing resources and the expertise to implement and utilize them. Most companies, aside from the largest, are excluded from accessing the highest levels of ML and AI.
Therefore, the next generation of advanced chips – which we might term “genius” chips – will prioritize efficiency and specialization. Chip designs will be optimized to suit the software they operate, resulting in significantly improved performance. When deploying powerful AI no longer necessitates an entire data center and becomes accessible to a wider range of businesses, the conditions for substantial innovation and disruption will be realized. Making expensive, resource-intensive AI more widely available is directly linked to these upcoming advancements in chip architecture and hardware design centered around software.
A refreshed emphasis on preparing for future innovations
The inherent characteristics of artificial intelligence present a unique difficulty for both those who develop and utilize AI-based hardware. The sheer scale of transformation is substantial: we are currently experiencing a transition from traditional coding practices performed by humans to software 2.0 – a realm where machine learning programs can be trained to ultimately operate autonomously. The speed of this transformation is also unparalleled; machine learning models can quickly become outdated, sometimes within a matter of months or weeks, and the techniques used for their training are continually being refined.
However, the development of new AI hardware still necessitates design, construction of prototypes, calibration, problem-solving, manufacturing, and delivery. The process from initial idea to a finalized product can span two years. While software has historically advanced at a faster pace than hardware, the current gap in speed is becoming insurmountable. We must adopt more innovative approaches to hardware design, anticipating a future that is increasingly difficult to foresee.
Indeed, the traditional concept of technology generations is starting to become less relevant. In the context of machine learning and AI, hardware must be engineered with the understanding that a significant portion of current knowledge will likely be outdated by the time the product is completed. Adaptability and the ability to be tailored will be the defining characteristics of successful hardware in the AI era, and this shift will likely benefit the entire marketplace.
Rather than concentrating investment on the latest model or a particular algorithm, organizations aiming to leverage these technologies will gain access to a wider range of processing systems that can adapt and change alongside the evolving requirements of machine learning and AI models.
This will empower businesses of all sizes and with varying levels of AI expertise to maintain their innovation and competitiveness for an extended period, and it will help avoid the stagnation that can occur when hardware restricts software capabilities – ultimately fostering more inventive and unforeseen AI applications across a broader range of organizations.
Growing acceptance of genuine AI and Machine Learning capabilities
I readily acknowledge the technology sector’s tendency to focus on the latest trends. There was a time when large datasets were considered the answer to all challenges, and the Internet of Things was predicted to revolutionize the world. Artificial intelligence has experienced similar cycles of inflated expectations (and arguably, has gone through them repeatedly). Currently, nearly every technology firm claims to utilize AI in some capacity, but often they are implementing relatively basic processes that are more accurately described as sophisticated data analysis.
I strongly believe that the transformative AI revolution we have anticipated has not yet materialized. However, within the next two to three years, it will begin to unfold as the infrastructure required for “true” AI capabilities becomes increasingly available. Forecasting the changes and disruptions resulting from broad access to the highest levels of potent machine learning and artificial intelligence is difficult—and that is precisely the key point!
Similar to how mobile phones empowered individuals globally, removing both technical and financial obstacles (generally speaking), the forthcoming generation of software-controlled hardware will be adaptable, configurable, and designed for longevity. The potential applications are limitless, and this will represent a significant shift in the technological landscape. The consequences of AI becoming more widely available and less expensive will extend beyond the technology industry, impacting numerous other sectors and fostering substantial innovation as powerful AI becomes accessible and affordable.
The substantial excitement surrounding AI—and the significant changes and advancements it was expected to deliver—will truly begin to take shape in the coming years. The underlying technology is currently under development or will soon be within reach of the many individuals across diverse industries who will leverage this new access to drive remarkable progress. We are particularly enthusiastic about contributing to this future and anticipate the advancements it will enable.