Der Blätterkatalog benötigt Javascript.
Bitte aktivieren Sie Javascript in Ihren Browser-Einstellungen.
The Blätterkatalog requires Javascript.
Please activate Javascript in your browser settings.
12 Elektronik International 2020 ARTIFICIAL INTELLIGENCE IMEC AND GLOBALFOUNDRIES ANALOG AI CHIP WITH UNDEMANDING NEEDS To get around elaborate data transfers between the Internet of Things and the cloud there are attempts to implement artificial intelligence at the IoT edge The analog AI chip AnIA with computing power of 2900 TOPS Wtakes us close to the aim By Gerhard Stelzer The Belgian research institute IMEC in Leuven and the semiconductor fab operator GlobalFoundries GF headquartered in Santa Clara California have cooperated to achieve a breakthrough in AI chips that enable deep neural networks in IoT edge devices The two partners have presented a hardware demonstrator chip with artificial intelligence This new chip is based on the AiMC analog inmemory computing architecture of IMEC and uses the 22FDX semiconductor fabrication process of GF The new chip is optimized for deep neural network calculations on inmemory computing hardware in the analog domain Achieving record high energy efficiency of up to 2900 TOPS Wteraoperations per second watt the AI accelerator is a key enabler for inferenceattheedge working in lowpower devices The advantages of this new technology when it comes to data privacy security and latency will impact AI applications in a wide range of de - vices from smart loudspeakers through to selfdriving vehicles Since the early days of digital computing the processor has been separated from memory Operations performed with a large amount of data require a similarly large number of data elements to be retrieved from memory This limitation called the von Neumann bottleneck can lengthen actual computing time especially in neural networks depending on large vector matrix multiplications Such computations are made with the precision of a digital computer and call for considerable energy But neural networks can produce results of comparable accuracy if vector matrix multiplications are performed with lower precision by analog technology Analog AI calculation with 2900 TOPS W Whereas the large vector matrix multiplications typical of AI tasks are performed bit by bit in the digital domain that works in the analog domain by Kirchhoffs rules with automatic current summing that allows for stored weights is how Maciej Wiatr process expert at GF explains the principle photo GlobalFoundries | IMEC