The Technical University of Munich (TUM) has announced a new breakthrough in AI chip technology that could revolutionize the industry. Professor Hussam Amrouch has developed an AI-ready architecture that boasts twice the power of comparable in-memory computing approaches, while consuming only half the energy. This development could have significant implications for generative AI, deep learning algorithms, and robotic applications.
Traditionally, chips were used solely for calculations, but Amrouch’s new approach integrates data storage into the transistors themselves, resulting in improved performance and energy efficiency. The transistors used in this architecture measure only 28 nanometers, with millions of them incorporated into each AI chip.
In order to meet the demands of future applications, AI chips need to be faster, more efficient, and less prone to overheating. Real-time calculations, such as those required during drone flights, can be extremely complex and energy-intensive for a computer. This is where the new AI chip excels. It is specifically designed to support such demanding tasks without compromising performance.
The success of AI chips is often measured by a parameter called TOPS/W, which stands for tera-operations per second per watt. Essentially, this parameter represents the number of trillion operations a processor can perform per second when given one watt of power. The new AI chip developed through the collaboration between Bosch, Fraunhofer IMPS, and GlobalFoundries is capable of delivering an impressive 885 TOPS/W. This puts it miles ahead of other comparable AI chips, including Samsung’s MRAM chip, which operates in the range of 10-20 TOPS/W. The development of such highly efficient chipsets opens up exciting possibilities for deep learning, generative AI, and robotics.
The concept behind this groundbreaking technology is inspired by the architecture of the human brain. Neurons process signals, while synapses store and recall information. In a similar vein, the new AI chip employs ferroelectric field effect transistors (FeFETs) that possess special characteristics, allowing them to store information even when disconnected from a power source. This enables the chip to store and process data simultaneously within its transistors.
While the potential of this new AI chip is immense, it is not expected to be commercially available for a few years. Professor Amrouch estimates that it will take three to five years, at the earliest, before in-memory chips suitable for real-world applications become widely accessible. This is due, in part, to the stringent security requirements of various industries. Before being adopted, a technology must not only function reliably but also meet the specific criteria of the sector it serves.
The achievement of Professor Amrouch and his team highlights the importance of interdisciplinary collaboration, bringing together researchers from computer science, informatics, and electrical engineering. Such collaborations foster innovative breakthroughs and drive progress in the field of hardware development. With the convergence of expertise, we can anticipate further advancements that will shape the future of AI and support the growth of various industries.
In conclusion, the development of this new AI chip represents a remarkable breakthrough that could transform the field of artificial intelligence. With its increased power and energy efficiency, the chip holds tremendous potential for applications such as deep learning, generative AI, and robotics. While it may take a few more years before these chips are widely available and meet industry-specific requirements, the progress made by Professor Amrouch and his team is a significant step forward. By pushing the boundaries of AI chip technology, they are paving the way for a future where complex tasks can be performed with exceptional efficiency and accuracy.
