Some of our readers were skeptical about the recent note that GlobalFoundries factories can become the forge of terminators, but in vain. GlobalFoundries is again in a hurry to surprise with the depth of interest in the topic of artificial intelligence and the production of silicon “brains”. This time together with Belgian developers.

Analog in Memory Computing (AiMC) IMEC neural network accelerator with decision making function

In a joint press release, Belgian research center Imec and GlobalFoundries announced a live demonstration of the new and unique AI chip. This processor or computational accelerator is designed based on the Analog in Memory Computing (AiMC) IMEC architecture and released using the GlobalFoundries 22nm process technology on FD-SOI (22FDX) wafers. Each of the companies has made its own contribution to what will soon become smart, autonomous and portable electronics.

The chip developed by Imec on the AiMC architecture (in Russian, analog computing in memory) has demonstrated a record high energy efficiency of calculations – up to 2900 TOPS (trillion operations per second) per watt. But this is not the limit. Imec promises to achieve an efficiency of 10,000 TOPS / W, which will make AI computing available to even the simplest battery-powered gadgets. And, of course, more sophisticated platforms such as self-driving cars, drones and robotics solutions will benefit from such chips.

The Analog in Memory Computing IMEC architecture bypasses a serious obstacle in the classical von Neumann logic – the so-called bottleneck of von Neumann bottleneck architecture. This limitation is due to the need to extract huge amounts of data from memory, which are then sent to the processor for processing. The data retrieval and transfer times can be much longer than the CPU processing time. This is especially critical for the operation of neural network accelerators, which rely on operations with multiplication of massive vector matrices, and this is all a lot of energy.

The AiMC IMEC architecture does all of the above differently. It performs calculations directly in the processor’s SRAM and does so not with digital data, but with data presented in analog form. “Analog” technologies make it possible to obtain practically the same result when multiplying vector matrices with the assumption of lower accuracy than when using data in the form of digital 0s and 1. Savings occur immediately on two points: by transferring data from memory to the processor and by the amount of data used for calculations. data.

As one of Imec’s machine learning leaders, Diederik Verkest, said, “Reference implementation [чипа] not only shows that analog computing in memory is possible in practice, but also that it achieves energy efficiency ten to one hundred times better than digital accelerators. “

GlobalFoundries plans to take this Imec development into service and in the future offer its customers as an option in the development and production of AI chips by interested parties. We add that the prototype chip was produced on 300-mm wafers at the GlobalFoundries plant in Dresden.

If you notice an error, select it with the mouse and press CTRL + ENTER.

By Alex

Alex Soojung-Kim Pang is a Silicon Valley-based consultant and writer. His latest book Rest: Why You Get More Done When You Work Less (Basic Books, 2016) and The Distraction Addiction (Little Brown, 2013) blend history, psychology, and neuroscience to explore the hidden role of leisure and mind-wandering in creative lives.

Leave a Reply

Your email address will not be published. Required fields are marked *