Written by: Varshith Uppalapati
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
The AI Revolution in Chip Architecture: Meeting the Demands of Tomorrow
The surge in demand for artificial intelligence (AI) applications has sparked a transformative wave in the semiconductor industry, propelling a shift from traditional general-purpose architectures to more specialized and application-specific designs. For those unaware, chip architecture encompasses the layout of a chip, encompassing factors such as size, shape, and component placement. It plays a pivotal role in dictating the overall performance of a chip, influencing crucial aspects including computation speed, power consumption, and versatility in supporting diverse applications. The challenges in current general-purpose architectures like x86 and ARM, despite their flexibility, lies in addressing the intricate requirements of AI and neural networks. Recognizing this limitation, startups and industry leaders are increasingly advocating for AI-specific designs capable of scaling across a wide range of power consumption levels – from the energy-efficient hundredths of a watt to the high-powered hundreds of watts. Due to the demands of AI applications, semiconductor manufacturers are turning towards application-specific integrated circuits (ASICs), renowned for their tailored designs optimized for specific tasks. Although the development of application specific ICs has become easier due to development in EDA tools, they present challenges such as non-recurring engineering (NRE) costs and the necessity for high production numbers. To address these challenges, the industry is contemplating a transition towards application-specific architectures that offer flexibility without compromising efficiency. It is estimated that ASICs will gain more ground against GPU and CPU, which are the current leaders in terms of market share, to meet the demands of deep learning. For inference, CPU, which currently accounts for 75% of the market, will lose its market share to ASICs as deep learning grows. As McKinsey estimates, CPU demand will drop to 50%, and ASIC will rise to 40% by 2025.
Moreover, semiconductor manufacturers are actively engaged in addressing the demands of AI applications through various approaches. While some are focused on the development of AI accelerators that seamlessly integrate with general-purpose cores, providing a performance boost while preserving versatility, others are pioneering entirely new architectures tailored specifically for AI tasks. By striking a balance between enhancing the efficiency of application-specific solutions and maintaining the adaptability of general-purpose architectures, this multi-faceted approach aims to meet the evolving needs of AI applications while fostering innovation and collaboration within the semiconductor industry.
Not Your Ordinary Microchip...
Red Semiconductor's innovation lies in its Vector Instruction Set Computing (VISC) architecture (Vector + RISC), born from the open-source hardware project Libre-SOC, which stands out as a potential game-changer in AI processing and is positioned to tackle the challenges associated with AI computations. Unlike traditional processors, VISC is designed to execute matrix mathematics using a vectorized instruction set. This approach aims to streamline various functions within a single architecture, eliminating the challenges posed by disparate languages and architectures that often accompany general-purpose solutions.
VISC's vectorised instruction sets automatic complex functions a predetermined number of times. This innovative approach minimises the need for frequent reading or writing to off-chip memory, addressing a common bottleneck in traditional architectures. The result is a significant enhancement in speed and power efficiency, particularly crucial for real-time execution in time-sensitive AI tasks. The emphasis on vectorisation is crucial for AI inference because it allows for efficient processing of large matrices and tensors commonly encountered in neural network computations. In AI inference, where rapid processing of data is essential for real-time or near-real-time decision-making, efficient vectorisation techniques like those employed by Red Semiconductor can significantly enhance performance. This efficiency not only accelerates inference tasks but also contributes to improved power efficiency, which is particularly important for edge devices and other resource-constrained environments.
Red Semiconductor's innovation targets edge applications, addressing key challenges like the performance gap in AI and cybersecurity threats. The company recently also signed a memorandum of understanding with Crypto Quantique, a leading provider of quantum-based security solutions for the IoT. This partnership further underscores Red Semiconductor's commitment to bolstering edge device security.
The Team
The start-up’s team boasts a wealth of industry experience and leadership. David Calderwood, Director of R&D and Chairman, has built teams at leading technology companies and served on the OpenPOWER Foundation's executive board.
James Lewis, CEO, is a prominent figure in the UK semiconductor scene, having founded successful ventures like Oxford Semiconductor and Redux. David Harold, COO, is renowned for his microprocessor marketing strategy expertise, previously holding the role of CMO at Imagination Technologies. Andrey Miroshnikov, Senior Engineer, brings valuable experience from his time at Qualcomm. Until now, the company has tapped funds from grants including the EU Horizon 2020 grant, but is now looking to raise £2.5m to finance its path towards full production.
The semiconductor industry is experiencing a transformative period fuelled by the AI boom. Various approaches, such as hardware accelerators, vector extensions, and innovative techniques like in-memory processing, are being explored to meet the escalating performance demands of AI and machine learning applications. While the UK government's £1 billion national semiconductor strategy marks a significant step forward, further support from the private sector may be essential to elevate the UK's standing in this field.
Comments