The video discusses the limitations of silicon transistors, which have been shrinking for decades, and the need for new materials and research into atomically thin materials. These materials, such as graphene and molybdenum disulfide, are promising due to their unique thermal and electrical properties. Researchers have developed a new device called an atomically thin memory sticker, which is a silicon wafer covered with thousands of these memory stickers. These memory stickers function like neurons in the brain, combining computation and data storage in one device.
From these atomic lithium memory stickers, researchers built an atomically thin microchip that performs memory computing. This research has attracted attention from top semiconductor companies and has been tested on a standard task of classifying handwritten digits, achieving 90% accuracy while consuming less power than a standard digital CMOS chip.
These new memory stickers, also called atom writers, are built in atomically thin materials and are used to implement spiking neural networks, which mimic the way our brain computes. This makes computation extremely efficient. The video also discusses the differences between conventional neural networks and spiking neural networks, and how the latter is more efficient for real-world applications. The researchers plan to fabricate 300 millimeter wafers in this technology and suggest that this is a combination of state-of-the-art materials, neuromorphic computing, and spiking neural networks.
1. Silicon transistors have been getting smaller for decades, approaching the limits at which they can no longer shrink to continue the Moore's Law [Source: Document 1].
2. Researchers are investigating new materials and atomically thin materials, which are promising [Source: Document 1].
3. Examples of atomically thin materials include graphene and molybdenum disulfide [Source: Document 1].
4. These materials are so thin that they are referred to as 2D materials [Source: Document 1].
5. Researchers have developed a new device, an atomically thin memory stick, which acts like neurons in our brain [Source: Document 1].
6. They built an atomically thin microchip, which performs memory computing [Source: Document 1].
7. This research has attracted attention from top semiconductor companies [Source: Document 1].
8. They created a spike neural network, which consists of thousands of these thin memory stickers [Source: Document 1].
9. They tested this chip on a standard task, classifying handwritten digits, and achieved 90% accuracy while consuming less power than a standard digital CMOS chip [Source: Document 1].
10. The new atom resters, also known as atom writers, are used to implement spike in neural networks [Source: Document 1].
11. Spike in neural networks mimic the way our brain computes, making this compute extremely efficient [Source: Document 1].
12. Spike in neural networks are rare fire spikes and shuffle less data than typical artificial neural networks [Source: Document 1].
13. They drain much less power, making spike in neural networks more efficient for real-world applications [Source: Document 1].
14. IBM previously experimented with 2D materials for chips, but eventually abandoned this approach [Source: Document 2].
15. In this work, they are using a slightly more robust material, a slightly thicker material [Source: Document 2].
16. The researchers used a compatible fabrication process to the one which is used for traditional chips [Source: Document 2].
17. They fabricated the 2D memory stair chip as a layer on top of the CMOS chip [Source: Document 2].
18. The CMOS chip is used to control the memory stairs and their switching [Source: Document 2].
19. Advanced atom registers can enable the chips of the future and new applications which previously didn't exist [Source: Document 2].
20. The researchers plan to fabricate 300 millimeter wafers in this technology [Source: Document 2].
21. This research was done in Polytechnico De Milano, one of the top Tech universities in Europe [Source: Document 3].